DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
(Submitted on 2/2/2026)
In regard to 103 rejections
- The applicant on Page 10 argues that the references do no teach the amended limitation “first statistical information that includes at least one first residual from fitting the first fitted label set that excludes both the first feature set, and first model information of the first machine model, from disclosure to the at least one module, wherein module is operative to fit, into at least one second fitted label set” for claims 1, 10 and 19. Specifically the applicant argument in regard to the reference “Tuv” that the claims require to adjust the target for the variables already selected through residuals and that the reference “Tuv” uses orthogonal predictors.
Examiner’s Response:
The examiner respectfully reject this argument. Perhaps known to a POSITA that Orthogonal predictors are uncorrelated variables in regression analysis that simplify model interpretation and reduce multicollinearity, enhancing the reliability of coefficient estimates. However, in a linear regression, the relationship between orthogonal predictors and residuals is a direct consequence of the orthogonality of residuals to the fitted space. For example, if X is orthogonal (columns are uncorrelated),
X
T
X is diagonal, and the OLS (Ordinary Least Squares) solution simplifies. The orthogonality of residuals still holds, and the residual sum of squares is simply the sum of squared projections onto the orthogonal complement of X. The examine reaffirms that “Tuv” is a valid reference. The examiner submits that the applicant argument is MOOT as a result of new references and new grounds of rejection. The examiner submits that better mapping for the new limitations new reference ”Zadeh” that teaches the requirement of not disclosing to the module. As the applicant has amended to focus on the module privacy the examiner interprets the module being a remote computing device based on the specification [0009 ]” in one example, this disclosure describes a method that includes: creating, by processing circuitry of a computing device, a learner unit by fitting, into a first fitted label set, an initial label set using at least one first learning technique, a machine learning model, and a first feature set; sending, by the processing circuitry of the computing device, to at least one module in a machine learning architecture, first statistical information defined by at least one first residual from fitting the first fitted label set, wherein the at least one module executes on at least one remote computing device”. Perhaps it known to a POSITA that a remote computing device that requires privacy protection is generally considered a user device in cybersecurity and policy contexts, even if it is not physically owned by the user. This is because it is used by a person to access or process sensitive data or systems, and its security posture directly affects the confidentiality, integrity, and compliance of that access. The examiner has used the new reference “ Zadeh” to teach the non-amended dependent claims as it is directly relevant to those claims.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-17 and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over
Hugh McMahan (hereinafter McMahan) US 2017/0109322 A1,
in view of Zili Li et.al. (hereinafter Li) US 11551083 B2,
in view of Lotfi Zadeh et.al. (hereinafter Zadeh) US 2018/0204111 A1.
In regard to claim 1: (Currently Amended)
McMahan discloses:
- A method comprising: creating, by processing circuitry of a computing device
In [0004]:
One example aspect of the present disclosure is directed to a computer-implemented method of updating a global model based on unevenly distributed data. The method includes receiving, by one or more computing devices, one or more local updates from a plurality of user devices. Each local update is determined by the respective user device based at least in part on one or more data examples stored on the respective user device. The one or more data examples stored on the plurality of user devices are distributed on an uneven basis, such that no user device includes a representative sample of an overall distribution of data examples. The method further includes aggregating, by the one or more computing devices
- wherein each of the at least one module executes on at least one remote computing device;
and wherein the computing device maintains the privacy in sending the first statistical information
in [0054]:
The technology discussed herein makes reference to servers, databases, software applications, and other computer-based systems, as well as actions taken and information sent to and from such systems.
In [0015]:
More particularly, in some embodiments, a set of input-output data can be used to describe a global objective via a loss function. Such functions can be, for instance, a convex or non-convex function, such as a linear regression function, logistic regression function. A local objective (F.sub.k) can also be defined using data stored on a computing device. For instance, the global objective can be defined as:
In [0030]:
FIG. 1 depicts an example system 300 for training one or more global machine learning models 306 using training data 308 stored locally on a plurality of user devices 302. System 300 can further include a server device 304. Server 304 can be configured to access machine learning model 306, and to provide model 306 to a plurality of user devices 302. Model 306 can be, for instance, a linear regression model,
In [0030]:
in some implementations, sever 304 can be configured to communicate with user devices 302 over one or more networks, such as network 240 of FIG. 3.
In [0052]:
The client device 230 can also include a network interface used to communicate with one or more remote computing devices (e.g. server 210) over the network 240.
in [0004]:
a computer-implemented method of updating a global model based on unevenly distributed data. The method includes receiving, by one or more computing devices, one or more local updates from a plurality of user devices.
in [0016]:
According to particular implementations of the present disclosure, the global objective can be solved by aggregating a plurality of local updates provided by a plurality of remote computing devices. Each remote computing device can, for instance, be a user device, such as a laptop computing device, desktop computing device, smartphone, tablet, wearable computing device, or other computing device. The local updates can be determined based at least in part on the respective local objectives.
in [0054]:
Databases and applications may be implemented on a single system or distributed across multiple systems.
in [0014]:
the local update may be determined using one or more gradient descent techniques
in [0014]:
the local update does not include the training data used to determine the local update. In this manner, the size of the local update can be independent of the training data used to determine the local update, thereby reducing bandwidth requirements and maintaining user privacy. In particular, a global model can be updated based at least in part on the received local updates. By only providing the local update (and not the training data) to the server, the global model update can be determined using reduced bandwidth requirements, and without compromising the security of potentially privacy sensitive data stored on the user devices
McMahan does not explicitly disclose:
- a learner unit by fitting, into a first fitted label set
- an initial label set using at least one first learning technique, a first machine learning model, and a first feature set;
- sending, by the processing circuitry of the computing device, to at least one module in a machine learning architecture, first statistical information that includes at least one first residual from fitting the first fitted label set that excludes both the first feature set, and first model information of the first machine model wherein module is operative to fit, into at least one second fitted label set
- receiving, by the processing circuitry of the computing device, and from the at least one module, second statistical information first statistical information that includes at least one second residual from fitting the second fitted label set that excludes both the second feature set, and second model information of the second machine model , and that is based on the first statistical information
- and updating, by the processing circuitry of the computing device, the learner unit by fitting, into a third fitted label set, the second statistical information using the at least one first learning technique and the first machine learning model.
However, Li discloses:
- a learner unit by fitting, into a first fitted label set
In [Col 4, lines 41-56]:
In accordance with various aspects of the present invention, the first configuration data includes hyperparameters for the neural network model and parameters for the first version of the neural network model. The hyperparameters may include one or more of: an architecture definition for the neural network model; a number of nodes for one or more layers in the neural network model; a set of node definitions indicating one or more of a node type and a node connectivity; a set of activation function definitions; and one or more cost function definitions. The parameters may include one or more of: weight values for one or more connections between nodes of the neural network model; weight values for one or more inputs to the neural network model; weight values for one or more recurrent paths in the neural network model; and bias values for one or more nodes of the neural network model.
In [ Col 9, lines 17-23]:
“neural network model” is used to refer to an artificial neural network that is configured to perform a particular data processing task. For example, in the case that a neural network model includes an acoustic model, the task may be to output phoneme or grapheme data (e.g. predictions of phonemes or graphemes) based on input audio data.
In [ Col 9, lines 26-32]:
In certain cases, a neural network model may be a model that is configured to provide a particular mapping between defined input data and defined output data. The input data may represent one modality and output data may represent another modality. The neural network model may be considered a function approximator that is trained on a set of data.
In [Col 19, lines 58-62]:
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570.
- an initial label set using at least one first learning technique, a first machine learning model, and a first feature set;
In [Col 12, lines 2-6]:
a version of the neural network model may be instantiated by implementing a class or class-like definition of the neural network model using initialization data. In this case, the initialization data includes the first configuration data 180.
- sending, by the processing circuitry of the computing device, to at least one module in a machine learning architecture,
In [Col 22, lines 49-53]:
a loss function may be a function of log it values across the ensemble of instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.)
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating the parameter data for the first version of the neural network model based on the second configuration data.
- receiving, by the processing circuitry of the computing device,
In [Col 22, lines 49-53]:
a loss function may be a function of log it values across the ensemble of instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.)
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating the parameter data for the first version of the neural network model based on the second configuration data.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, and Li.
McMahan teaches providing user privacy.
Li teaches learner unit.
One of ordinary skill would have motivation to combine McMahan and Li that can improve the versions of neural network model without exchanging any private information (Li [Col 2, lines 54-66])
McMahan and Li do not explicitly disclose:
- first statistical information that includes at least one first residual from fitting the first fitted label set that excludes both the first feature set, and first model information of the first machine model, from disclosure to the at least one module, wherein module is operative to fit, into at least one second fitted label set
- second statistical information first statistical information that includes at least one second residual from fitting the second fitted label set that excludes both the second feature set, and second model information of the second machine model, and that is based on the first statistical information, from disclosure to the at least one module, and that is based on the first statistical information;
- and updating, by the processing circuitry of the computing device, the learner unit by fitting into a third fitted label set, the second statistical information using the at least one first learning technique and the first machine learning model.
However, Zadeh discloses:
- first statistical information that includes at least one first residual from fitting the first fitted label set that excludes both the first feature set, and first model information of the first machine model, from disclosure to the at least one module, wherein module is operative to fit, into at least one second fitted label set
[3162]:
the learning process of the learning machine uses a statistical approach to hypothesis parameters by keeping track of those values over time during fitting the data fitting, with more weight given to those parameters occurring more often.
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership, the set is the fitted labeled set)
[2129]:
For the aggregation method (also called ensemble learning, or boosting, or mixture of experts), we have a learning which tries to replicate the function independently (not jointly), and then combine and put them together later,
[2129]:
For the aggregation method, for regression or real number cases, we take an average or weighted average,
[2129]:
For the aggregation method, we have 2 types: (a) After-the-fact situation (where we already have the solutions, and then we combine them), and (b) Before-the-fact situation (where we get solutions, with the view or intention or assumption to blend or combine them together later). For the aggregation method, as one example, we have the Boosting method, where we enforce the decorrelation (not by chance), e.g. by building one hypothesis at a time, for a good mixture, sequentially.
[2092] :
In one embodiment, for aggregating the correlated fuzzy sets, for the Min-Max aggregation method, we can get the final membership value, μ.sub.final, based on the individual membership values, μ.sub.i and μ.sub.2, as (where index i runs from 0 to n):
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
[1638]:
An approach to approximate or render the overlap (15) between the category sets, such as C.sub.X,n, may use α-cuts to present each crisp α-cuts of predetermined category set as a set of points in (m,σ) space. These sets of points may be modeled efficiently, e.g., based on graphical models, optimized for fast transformation and intersection operations. For example, the models that use peripheral description for the α-cuts allow robust and efficient determination of intersection and avoid the need to transform all the points within the set individually, in order to reduce the computation involved in (13).
(BRI: Elements and their degree of membership and within the context of A and B as membership. As there is overlapping between categories sets (fuzzy sets), it does represent plurality of fitted label sets because an element can have multiple non-zero membership degrees, each indicating the extent to which it fits into a different categories)
[1723]:
In one embodiment, the added units and the previous units are used to make association and/or correlation with labeled samples, e.g., during the supervised training.
[1723]:
In one embodiment, the labels are continuous valued (or multi-valued), e.g., having values in range [0, 1], to indicate the degree in which the document is classified by a label (or the membership function of the document in the label's class).
[0113]:
For one embodiment: Decisions are based on information. To be useful, information must be reliable. Basically, the concept of a Z-number relates to the issue of reliability of information. A Z-number, Z, has two components, Z=(A,B). The first component, A, is a restriction (constraint) on the values which a real-valued uncertain variable, X, is allowed to take. The second component, B, is a measure of reliability (certainty) of the first component.
[0564]:
For purposes of computation, when A and B are described in a natural language, the meaning of A and B is precisiated (graduated) through association with membership functions, μ.sub.A and μ.sub.B, respectively, FIG. 1.
[0593]:
A basic fuzzy if-then rule may be expressed as: if X is A then Y is B, where A and B are fuzzy numbers. The meaning of such a rule is defined as:
PNG
media_image2.png
17
202
media_image2.png
Greyscale
[1678] :
In one embodiment, the learning is achieved through simulation using a data (and label) sample generation based on one or more models. In one embodiment, a network trained based on model(s) is used to recognize and classify actual data which may not have been seen before. In one embodiment, the system is trained to infer the potential model(s) itself by recognizing the (e.g., observed) data conforming to a particular model and its associated labels/parameters.
[1757]:
The features of both images (and in particular the differences between their features) are correlated using a correlator/analyzer module (e.g., having unit/neurons) with the label/feature difference identified in the label layer (e.g., L.sub.1). In one embodiment, the L layer represents the labels indicated the feature differences between the images (denoted by ΔL). In one embodiment, more than one label is selected by the controller (indicating the differences between the features of the images selected from the database for training the correlation).
[2431]:
In one embodiment, the system uses statistics and patterns extracted from images
[2788]:
a script is executed that iterate through the document object model or the window object to identify elements associated with images
[1723]:
In one embodiment, feature detection system is used to train document classification based on learned (e.g., unsupervised) features corresponding to documents based on terms contained in the document (such as statistics of several hundred or several thousand common words)
[2521]:
In one embodiment, as for example depicted in FIG. 228, based on an identifier or a URL, a movie and its associated data (e.g., metadata, title, description, owner/uploader, channel, comments, likes, and statistics) are extracted from a repository. In one embodiment, the movie and its associated data are analyzed, e.g., by a video/audio analyzer and keyword/concept extraction/analyzer, to provide/generate features/annotation and metadata
[1757]:
In one embodiment, the weights related to the units in the correlator/analyzer are trained to detect the feature differences by a stochastic or batch learning algorithm.
[1734]:
features of an object (e.g., pose including rotation) is determined, and based on such features, features of sub-objects of other objects depicted in an image are extracted by preprocessing (e.g., mapping) a portion of an image into a segmented layout with variable resolution. Then, the mapped image (or portion thereof) is provided to a classifier or feature recognition system to determine the features from the mapped image
1351]:
In one embodiment, an optimum statistical classifier is used. In one embodiment, a Bayes classifier is used
[1351]:
one embodiment, a perceptron for 2-pattern classes is used. In one embodiment, the least mean square (LMS) delta rule for training perceptrons is used, to minimize the error between the actual response and the desired response (for the training purposes). FIG. 115 is an example of a system described above.
[1352] :
In one embodiment, a multi-layer feed-forward neural network is used. In one embodiment, the training is done by back propagation, using the total squared error between the actual responses and desired responses for the nodes in the output layer.
[2301]:
In one embodiment, we segment the data of any type, including video, sound, and multimedia, based on sudden change in the sequence (or big delta or difference),
[2300]:
In one embodiment, we use Bayesian model, for both sides of the potential boundary between segments, with 2 different model parameters, to fit the 2 sides better, to examine the potential boundary for segmentation, e.g. for speech.
(BRI: in the context of a neural network or any predictive model, the difference between the actual output (response) and the desired output (target) for the nodes in the output layer is indeed the residual. From a statistical information perspective, these residuals are not just raw differences and they are statistical information in the sense that they quantify the discrepancy between the model’s predictions and the true data. The plurality of residuals are within the context of residual associated to each node and the statistical information associated to each residual)
[1768] :
In one embodiment, meta data such as the GPS data (or for example other accompanying metadata captured with images taken from mobile devices such as smart phones) are used as labels (e.g., continuous valued).
[1586]:
The communication between different units, devices, or modules are done by wire, cable, fiber optics, wirelessly, WiFi, Bluetooth, through network, Internet, copper interconnect, antenna, satellite dish, or the like.
1759]:
In one embodiment, one or more pose detection modules (e.g., based on edge detection or color region/shape) are used to determine the pose of a face within an image/data
[1810]:
FIG. 127 shows a system for context determination, with language input device, which feeds dissecting and parsing modules to get the components or parts of the sentence, which feeds the analyzing module (which e.g. may include memory units and processor units or CPU or computing module), which is connected to the context determination module, which is connected to the default analyzer module and multiple other context analyzer module
PNG
media_image3.png
537
513
media_image3.png
Greyscale
[2422]:
In one embodiment, we have private or public or semi-public or semi-private (or the like) settings for our sharing or displaying or reviewing or tagging or annotating or accessing or searching or browsing of images or objects or videos, for user, friends, family, co-worker, boss, employee, contractor, senior management, public, social network, college, school, classmate, roommate, household, shared device, shared account, friend-of-friend, friend-of-friend-of-friend, and so on, or the like. In one embodiment, we have government excluding list database, for specified individuals, to exclude for the rules, for the above functions, for privacy settings. In one embodiment, we have the intersection of the privacy settings of multiple users or contributors
[2789]:
In one embodiment, the script determines whether to identify the images based on the domain name (e.g., as positive filter to include or negative filter to exclude).
[2830]:
publisher's restriction/filters to exclude certain or certain types of merchants and merchant's restriction/filters to exclude certain or certain types of publishers.
[1917]:
In one embodiment, for conditional relationships, or multiple choices, we can continue, until we get to a dead end or conflict, and then, backtrack to eliminate or adjust one or more choices, on the chain going backward, to correct or adjust some assumptions, choices, or conditions, on the way.
[2355]:
we can input these conditions or rules into our rule engine, or use it for prediction, control system, forecasting (economy, elections, and other events), social behavioral analysis, consumer behavioral analysis, predicting revolutions or unrest, detecting frauds, detecting unusual behaviors, detecting unusual patterns, finding liars or contradictions
- second statistical information first statistical information that includes at least one second residual from fitting the second fitted label set that excludes both the second feature set, and second model information of the second machine model, and that is based on the first statistical information, from disclosure to the at least one module, and that is based on the first statistical information;
[3162]:
the learning process of the learning machine uses a statistical approach to hypothesis parameters by keeping track of those values over time during fitting the data fitting, with more weight given to those parameters occurring more often.
[2129]:
For the aggregation method (also called ensemble learning, or boosting, or mixture of experts), we have a learning which tries to replicate the function independently (not jointly), and then combine and put them together later,
[2129]:
For the aggregation method, for regression or real number cases, we take an average or weighted average,
[2129]:
For the aggregation method, we have 2 types: (a) After-the-fact situation (where we already have the solutions, and then we combine them), and (b) Before-the-fact situation (where we get solutions, with the view or intention or assumption to blend or combine them together later). For the aggregation method, as one example, we have the Boosting method, where we enforce the decorrelation (not by chance), e.g. by building one hypothesis at a time, for a good mixture, sequentially.
[2092] :
In one embodiment, for aggregating the correlated fuzzy sets, for the Min-Max aggregation method, we can get the final membership value, μ.sub.final, based on the individual membership values, μ.sub.i and μ.sub.2, as (where index i runs from 0 to n):
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
[0565]:
The membership function of A, μ.sub.A, may be elicited by asking a succession of questions of the form: To what degree does the number, a, fit your perception of A? Example: To what degree does 50 minutes fit your perception of about 45 minutes? The same applies to B. The fuzzy set, A, may be interpreted as the possibility distribution of X. The concept of a Z-number may be generalized in various ways. In particular, X may be assumed to take values in R.sup.n, in which case A is a Cartesian product of fuzzy numbers.
[1638]:
An approach to approximate or render the overlap (15) between the category sets, such as C.sub.X,n, may use α-cuts to present each crisp α-cuts of predetermined category set as a set of points in (m,σ) space. These sets of points may be modeled efficiently, e.g., based on graphical models, optimized for fast transformation and intersection operations. For example, the models that use peripheral description for the α-cuts allow robust and efficient determination of intersection and avoid the need to transform all the points within the set individually, in order to reduce the computation involved in (13).
(BRI: Elements and their degree of membership and within the context of A and B as membership. As there is overlapping between categories sets (fuzzy sets), it does represent plurality of fitted label sets because an element can have multiple non-zero membership degrees, each indicating the extent to which it fits into a different categories)
[1638]:
An approach to approximate or render the overlap (15) between the category sets, such as C.sub.X,n, may use α-cuts to present each crisp α-cuts of predetermined category set as a set of points in (m,σ) space. These sets of points may be modeled efficiently, e.g., based on graphical models, optimized for fast transformation and intersection operations. For example, the models that use peripheral description for the α-cuts allow robust and efficient determination of intersection and avoid the need to transform all the points within the set individually, in order to reduce the computation involved in (13).
[1723]:
In one embodiment, the added units and the previous units are used to make association and/or correlation with labeled samples, e.g., during the supervised training.
[1723]:
In one embodiment, the labels are continuous valued (or multi-valued), e.g., having values in range [0, 1], to indicate the degree in which the document is classified by a label (or the membership function of the document in the label's class).
[0113]:
For one embodiment: Decisions are based on information. To be useful, information must be reliable. Basically, the concept of a Z-number relates to the issue of reliability of information. A Z-number, Z, has two components, Z=(A,B). The first component, A, is a restriction (constraint) on the values which a real-valued uncertain variable, X, is allowed to take. The second component, B, is a measure of reliability (certainty) of the first component.
[0564]:
For purposes of computation, when A and B are described in a natural language, the meaning of A and B is precisiated (graduated) through association with membership functions, μ.sub.A and μ.sub.B, respectively, FIG. 1.
[0593]:
A basic fuzzy if-then rule may be expressed as: if X is A then Y is B, where A and B are fuzzy numbers. The meaning of such a rule is defined as:
PNG
media_image2.png
17
202
media_image2.png
Greyscale
[1678] :
In one embodiment, the learning is achieved through simulation using a data (and label) sample generation based on one or more models. In one embodiment, a network trained based on model(s) is used to recognize and classify actual data which may not have been seen before. In one embodiment, the system is trained to infer the potential model(s) itself by recognizing the (e.g., observed) data conforming to a particular model and its associated labels/parameters.
[1757]:
The features of both images (and in particular the differences between their features) are correlated using a correlator/analyzer module (e.g., having unit/neurons) with the label/feature difference identified in the label layer (e.g., L.sub.1). In one embodiment, the L layer represents the labels indicated the feature differences between the images (denoted by ΔL). In one embodiment, more than one label is selected by the controller (indicating the differences between the features of the images selected from the database for training the correlation).
[2431]:
In one embodiment, the system uses statistics and patterns extracted from images
[2788]:
a script is executed that iterate through the document object model or the window object to identify elements associated with images
[1723]:
In one embodiment, feature detection system is used to train document classification based on learned (e.g., unsupervised) features corresponding to documents based on terms contained in the document (such as statistics of several hundred or several thousand common words)
[2521]:
In one embodiment, as for example depicted in FIG. 228, based on an identifier or a URL, a movie and its associated data (e.g., metadata, title, description, owner/uploader, channel, comments, likes, and statistics) are extracted from a repository. In one embodiment, the movie and its associated data are analyzed, e.g., by a video/audio analyzer and keyword/concept extraction/analyzer, to provide/generate features/annotation and metadata
[1757]:
In one embodiment, the weights related to the units in the correlator/analyzer are trained to detect the feature differences by a stochastic or batch learning algorithm.
[1734]:
features of an object (e.g., pose including rotation) is determined, and based on such features, features of sub-objects of other objects depicted in an image are extracted by preprocessing (e.g., mapping) a portion of an image into a segmented layout with variable resolution. Then, the mapped image (or portion thereof) is provided to a classifier or feature recognition system to determine the features from the mapped image
[1734]:
in one embodiment, based on the perspective/skew/projection of the frame (or other indicators), the image or a portion of image is mapped to a segmented layout for input to a network for further feature detection or classification.
[1351]:
In one embodiment, an optimum statistical classifier is used. In one embodiment, a Bayes classifier is used
[1351]:
one embodiment, a perceptron for 2-pattern classes is used. In one embodiment, the least mean square (LMS) delta rule for training perceptrons is used, to minimize the error between the actual response and the desired response (for the training purposes). FIG. 115 is an example of a system described above.
[1352] :
In one embodiment, a multi-layer feed-forward neural network is used. In one embodiment, the training is done by back propagation, using the total squared error between the actual responses and desired responses for the nodes in the output layer.
[2301]:
In one embodiment, we segment the data of any type, including video, sound, and multimedia, based on sudden change in the sequence (or big delta or difference),
[2300]:
In one embodiment, we use Bayesian model, for both sides of the potential boundary between segments, with 2 different model parameters, to fit the 2 sides better, to examine the potential boundary for segmentation, e.g. for speech.
(BRI: in the context of a neural network or any predictive model, the difference between the actual output (response) and the desired output (target) for the nodes in the output layer is indeed the residual. From a statistical information perspective, these residuals are not just raw differences and they are statistical information in the sense that they quantify the discrepancy between the model’s predictions and the true data. The plurality of residuals are within the context of residual associated to each node and the statistical information associated to each residual)
[1768] :
In one embodiment, meta data such as the GPS data (or for example other accompanying metadata captured with images taken from mobile devices such as smart phones) are used as labels (e.g., continuous valued).
[1586]:
The communication between different units, devices, or modules are done by wire, cable, fiber optics, wirelessly, WiFi, Bluetooth, through network, Internet, copper interconnect, antenna, satellite dish, or the like.
1759]:
In one embodiment, one or more pose detection modules (e.g., based on edge detection or color region/shape) are used to determine the pose of a face within an image/data
[1810]:
FIG. 127 shows a system for context determination, with language input device, which feeds dissecting and parsing modules to get the components or parts of the sentence, which feeds the analyzing module (which e.g. may include memory units and processor units or CPU or computing module), which is connected to the context determination module, which is connected to the default analyzer module and multiple other context analyzer module
PNG
media_image3.png
537
513
media_image3.png
Greyscale
[2422]:
In one embodiment, we have private or public or semi-public or semi-private (or the like) settings for our sharing or displaying or reviewing or tagging or annotating or accessing or searching or browsing of images or objects or videos, for user, friends, family, co-worker, boss, employee, contractor, senior management, public, social network, college, school, classmate, roommate, household, shared device, shared account, friend-of-friend, friend-of-friend-of-friend, and so on, or the like. In one embodiment, we have government excluding list database, for specified individuals, to exclude for the rules, for the above functions, for privacy settings. In one embodiment, we have the intersection of the privacy settings of multiple users or contributors
[2789]:
In one embodiment, the script determines whether to identify the images based on the domain name (e.g., as positive filter to include or negative filter to exclude).
[2830]:
publisher's restriction/filters to exclude certain or certain types of merchants and merchant's restriction/filters to exclude certain or certain types of publishers.
[1917]:
In one embodiment, for conditional relationships, or multiple choices, we can continue, until we get to a dead end or conflict, and then, backtrack to eliminate or adjust one or more choices, on the chain going backward, to correct or adjust some assumptions, choices, or conditions, on the way.
[2355]:
we can input these conditions or rules into our rule engine, or use it for prediction, control system, forecasting (economy, elections, and other events), social behavioral analysis, consumer behavioral analysis, predicting revolutions or unrest, detecting frauds, detecting unusual behaviors, detecting unusual patterns, finding liars or contradictions
- and updating, by the processing circuitry of the computing device, the learner unit by fitting into a third fitted label set, the second statistical information using the at least one first learning technique and the first machine learning model into a third fitted label set, the second statistical information using the at least one first learning technique and the first machine learning model.
[1340]:
the fuzzy classifier module or device classifies or separates different pictures into clusters or groups in N-dimensional feature space.
[1531]:
a method for fuzzy logic control, in which an input module receives a precisiated proposition associated with a protoform. A fuzzy logic inference engine evaluates a first fuzzy logic rule from the fuzzy logic rule repository. The fuzzy logic inference engine is in or loaded on or executed on or implemented in a computing device, which comprises one or more of following: computer, processor device
[0565]:
The membership function of A, μ.sub.A, may be elicited by asking a succession of questions of the form: To what degree does the number, a, fit your perception of A? Example: To what degree does 50 minutes fit your perception of about 45 minutes? The same applies to B. The fuzzy set, A, may be interpreted as the possibility distribution of X. The concept of a Z-number may be generalized in various ways. In particular, X may be assumed to take values in R.sup.n, in which case A is a Cartesian product of fuzzy numbers. Simple examples of Z-valuations are:
(BRI: When fitting a model to data, the goal is to find parameters that best match observed values. With fuzzy data, the “fit” is adjusted to account for uncertainty)
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership, the set is the fitted labeled set)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
1346]).
In regard to claim 2: (Original)
McMahan, and Li do not explicitly disclose:
- generating, from a new feature set and the learner unit, a first set of predicted labels.
However, Zadeh discloses:
- generating, from a new feature set and the learner unit, a first set of predicted labels.
[3022]:
detecting or classifying a feature set from said image;
[3023]:
taking an optimization step in training a correlation layer using said feature set and one or more of said invariant or semi-invariant parameters, said variant parameters, and said pose parameters, as input to said correlation layer; and
[3029]:
taking an optimization step in training a correlation layer using said first feature set and said second feature set as input to said correlation layer;
[3031]:
wherein said correlation layer, upon training, outputs a translated feature set, given a third feature set as input to said correlation layer.
(BRI: translated feature set is the new feature)
3169]:
One embodiment uses predictive feature detection. In one embodiment, the inference module predicts where the features might be based on the initial recognition.
[2782]:
in one embodiment, the indexing uses fuzzy values and intervals. In one embodiment, the search provides one or more potential matching results to a match maker module. In one embodiment, the feature values (fuzzy, crisp, labels) are also provided for coded features (i.e., the features that are descriptive).
[2533]:
In one embodiment, a morph module/application is used to make adjustment to descriptive features/labels after recognizing the model features.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
In regard to claim 3: (Original)
McMahan, and Li do not explicitly disclose:
- querying the at least one module for a second set of predicted labels for the new feature
However, Zadeh discloses:
- querying the at least one module for a second set of predicted labels for the new feature
[2994]:
an input module receiving a first data;
[2983]:
said relevance analysis module generating one or more second relevant items from said one or more first relevant items;
(BRI: a first input module receiving a first item, and a relevance analysis module generating one or more second relevant items from the first relevant items does represent a query in the sense of information retrieval and filtering)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
In regard to claim 4: (Original)
McMahan, and Li do not explicitly disclose:
- combining the first set of predicted labels and the second set of predicted labels into a final set of predicted labels.
[2770]:
The features representing the object/person/face are combined or enhanced based on the reliability of the features from different sources.
[1739]:
In one embodiment, the locations of interest (e.g., the location of faces within an image) is determined by a scanning the image through a variable size window over an image at different location on the image, searching for example for particular features or signatures (e.g., head or face).
[1750]:
In one embodiment, classifiers are trained to detect high level signatures/features of various objects/concepts, e.g., by training the classifiers with (labeled) training data sets,
[1739]:
such image and the associated information are used to train a feature detector/classifier to learn or predict the focuses of interest, by correlating/associating the image features with the locations of interest. In one embodiment, the image and various positions of interest are iteratively inputted to the system during training. The stochastic nature of the correlation layer, stochastically reconstruct parameters associated with the location of interest as output, e.g., using an RBM.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
In regard to claim 5: (Original)
McMahan, and Li do not explicitly disclose:
- repeating the sending and the receiving until an out-sample error satisfies a criterion, wherein the out-sample error is computed by cross-validation.
However, Zadeh discloses:
- repeating the sending and the receiving until an out-sample error satisfies a criterion, wherein the out-sample error is computed by cross-validation.
[1781]:
In one embodiment, linear models, such as perceptron, linear regression, and/or logistic regression are used. For example, perceptron is used for classification, e.g., in or out, accept or deny, and is trained by perceptron learning algorithm including a pocket version. The linear regression is for example used to determine (continuous valued or real valued) amount, based on squared error function and pseudo-inverse algorithm.
[1781]:
VC (Vapnik-Chervonenkis) dimension for a Hypothesis set (i.e., he most points that can be shuttered by the hypothesis set) is related to hypothesis set's growth function, and in one embodiment, the VC inequality (in terms of growth function and number of training samples) provides a rule of experience for the number of points needed for training (e.g., >10×VC dimension). The VC inequality places an upper bound on the probability of the out-of-sample error (i.e., the generalization error) is within the in-sample error by a given error margin and a given number of in-sample (training) data.
[1781]:
One embodiment uses an adaptive learning rate. In one embodiment, the default learning rate is 0.1. In one embodiment, the number of iterations of epoch is limited to a maximum (early stopping), in order to avoid over fitting the noise/error and deteriorate generalization by increasing the out of sample error.
[0320] :
FIG. 56 shows how to build a fuzzy model, going through iterations, to validate a model, based on some thresholds or conditions.
[2885]:
break down a larger QD problem into series of smaller ones, or breaking the problem to the smallest chunk in pair-wise sequential minimal optimization. In addition, the solver repeats the optimization by varying the values of C and/or kernel parameter(s) within a wide exponential range, and a grid search is used to determine the optimum hyperparameter(s) likely to minimize out of sample error (e.g., estimated by validation dataset).
[2045]:
In one embodiment, we recognize faces in the album and insert it automatically in the email, for sender or receiver, so that it would be easier to recognize the people in the email list
[2451]:
in one embodiment, the user sends the photo to analyzer for analysis and recognition, so that the name and other information for the person in photo are obtained and returned back to the use
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
In regard to claim 6: (Original)
McMahan does not explicitly disclose:
- the learner unit and the at least one module implement aligned or partially aligned feature datasets.
However, Li discloses:
the learner unit and the at least one module implement aligned or partially aligned feature datasets.
In [Col 23, lines 31-35]:
In accordance with an embodiment and aspect of the invention where gradient data is exchanged, differences in processing rate may cause differences in the amount of data that is evaluated at each device, which may in turn contribute to uncertainty in the gradient calculations.
In [Col 23, lines 41-50]:
In this case, the gradient data may be aggregated with a combining function that weights the different sets of gradient data inversely to a value of the uncertainty metric. In certain cases, the master device may be configured to determine a “samples per second” data processing rate on one or more slave devices. The master device may then dynamically adjust the number of samples per iteration on one or more of the master device and the slave devices to align iteration synchronization with the master device.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, and Li.
McMahan teaches providing user privacy.
Li teaches learning unit.
One of ordinary skill would have motivation to combine McMahan and Li that can improve the versions of neural network model without exchanging any private information (Li [Col 2, lines 54-66])
In regard to claim 7: (Original)
McMahan does not explicitly disclose:
- selecting the at least one module to run in the machine learning architecture based on at least one of communication bandwidth, cost constraints, or computational overhead.
However, Li discloses:
- selecting the at least one module to run in the machine learning architecture based on at least one of communication bandwidth, cost constraints, or computational overhead.
in [Col 23, lines 20-30]:
In accordance with an embodiment and aspect of the invention gradient data is exchanged, due to differences between processing speeds on the slave devices and the master device and/or due to differences in the amount of data that each has available for training, there may be large differences in processing time per iteration of configuration data exchange (or per epoch of training). A specific target cycle time may be set (e.g. as a predefined interval) and a back-propagation algorithm on one or more of the master and slave devices may be configured to process enough training data to meet the target cycle time.
In [Col 26, lines 44-49]:
In certain examples, the hyperparameters of the neural network model may be defined according to a predefined standard, and the binary executable may be designed specifically for this standard. This may help to minimize data exchange between the master device and the slave devices, which may speed training iterations.
In [Col 26, lines 39-41]:
Data privacy may also be verified by checking the bandwidth of network traffic between the master device and the slave device.
In [Col 22, lines 63-67], In [Col 23, lines 1-6]:
the master device may be configured to use aggregate data derived from the second configuration data output by the plurality of slave devices to update parameters for the first version of the neural network model. For example, gradient data from the plurality of slave devices may be aggregated by averaging or another suitable statistical function. In one case, the gradient data from different slave devices may be compared to selectively update the parameters for the first version of the neural network model.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, and Li.
McMahan teaches providing user privacy.
Li teaches learning unit.
One of ordinary skill would have motivation to combine McMahan and Li that can improve the versions of neural network model without exchanging any private information (Li [Col 2, lines 54-66])
In regard to claim 8: (Previously Presented)
McMahan does not explicitly disclose:
- creating, by the processing circuitry of the computing device, the learner unit further comprises training the first machine learning model using the at least one first learning technique with the initial label set and the first feature set,
- wherein the trained machine learning model is configured to generate the first fitted label set for the first feature set, the first statistical information further comprises determining a first particular residual of the at least one first residual based on a first fitted label of the first fitted label set and at least one of a first initial label of the initial label or first observed data in the first feature set, wherein the at least one module trains the least one second machine learning model using the at least one second learning technique with the at least one second feature set,
- wherein a second particular residual of the least one second residual is determined based on a second fitted label of the second fitted label set and at least one of the first particular residual, the first initial label of the initial label set, or first observed data in the second feature set, and
- wherein updating, by the processing circuitry of the computing device, the learner unit by fitting, into the third fitted label set, the second statistical information further comprises further training the trained machine learning model with the at least one second residual and the first feature set;
- third statistical information further defined by at least one third residual from fitting the third fitted label set.
- wherein sending, by the processing circuitry of the computing device [[,]], and to the at least one module in the machine learning architecture,
- and further comprising sending, by the processing circuitry of the computing device, to the at least one module in the machine learning architecture,
However, Li discloses:
- creating, by the processing circuitry of the computing device, the learner unit further comprises training the first machine learning model using the at least one first learning technique with the initial label set and the first feature set,
In [Col 19, lines 58-62]:
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570.
- wherein the trained machine learning model is configured to generate the first fitted label set for the first feature set, the first statistical information further comprises determining a first particular residual of the at least one first residual based on a first fitted label of the first fitted label set and at least one of a first initial label of the initial label or first observed data in the first feature set, wherein the at least one module trains the least one second machine learning model using the at least one second learning technique with the at least one second feature set,
In [Col 6, lines 47-55]:
In accordance with various aspects of the present invention, the receiving includes receiving second configuration data from a plurality of slave devices, and the updating includes: instantiating an ensemble of second versions of the neural network model as an ensemble of teacher models, and using the ensemble of teacher models to train the student model. The first version of the neural network model may thus be improved using private data from a large variety of heterogeneous devices.
The slave devices 722, 724, and 726 are communicatively coupled to the master device 710 via one or more communication networks 750. The master device 710 is configured to transmit first configuration data 780 for distribution to the plurality of slave devices 722, 724, and 726 via one or more communication networks 750.
In [Col 15, lines 15-19]:
Referring now to FIG. 2, a distributed training system 200 is shown in accordance with an embodiment of the invention. The distributed training system 200 may be seen as a variation of the first distributed training system 100 of FIG. 1 and similar features share similar reference numerals
In [Col 22, lines 48-63]:
a loss function may be a function of log it values across the ensemble of
instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.). Alternatively, the first version of the neural network model may be trained iteratively with each of the instantiated second versions being used as teachers in turn. In certain examples, the master device is able to evaluate and exclude slave configuration data from its teacher ensemble based on a predefined level of performance on a validation dataset, e.g. if an instantiated second version of the neural network model has an accuracy score below a threshold on the validation dataset, it may be excluded from the ensemble.
In [Col 17, lines 41-52]:
the first and second versions of the neural network model 360 and 370 start as copies of a common neural network configuration, e.g. that is initialized based on the first configuration data 380. Thus, just the parameters of the second version of the neural network model 370 change during training. The updated parameters is stored in the data storage device 372. An output from the first version of the neural network model 360 forms part of a loss function for the second version of the neural network model 370, such that during training the second version of the neural network model 370 “learns” to approximate the first version of the neural network model 360.
In [Col 19, lines 58-67], In [Col 20, lines 1-3] :
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570
(BRI: the plurality of fitted data is within the context of training plurality of slave devices)
- wherein a second particular residual of the least one second residual is determined based on a second fitted label of the second fitted label set and at least one of the first particular residual, the first initial label of the initial label set, or first observed data in the second feature set, and
In [Col 6, lines 47-55]:
In accordance with various aspects of the present invention, the receiving includes receiving second configuration data from a plurality of slave devices, and the updating includes: instantiating an ensemble of second versions of the neural network model as an ensemble of teacher models, and using the ensemble of teacher models to train the student model. The first version of the neural network model may thus be improved using private data from a large variety of heterogeneous devices.
The slave devices 722, 724, and 726 are communicatively coupled to the master device 710 via one or more communication networks 750. The master device 710 is configured to transmit first configuration data 780 for distribution to the plurality of slave devices 722, 724, and 726 via one or more communication networks 750.
In [Col 15, lines 15-19]:
Referring now to FIG. 2, a distributed training system 200 is shown in accordance with an embodiment of the invention. The distributed training system 200 may be seen as a variation of the first distributed training system 100 of FIG. 1 and similar features share similar reference numerals
In [Col 22, lines 48-63]:
a loss function may be a function of log it values across the ensemble of
instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.). Alternatively, the first version of the neural network model may be trained iteratively with each of the instantiated second versions being used as teachers in turn. In certain examples, the master device is able to evaluate and exclude slave configuration data from its teacher ensemble based on a predefined level of performance on a validation dataset, e.g. if an instantiated second version of the neural network model has an accuracy score below a threshold on the validation dataset, it may be excluded from the ensemble.
In [Col 17, lines 41-52]:
the first and second versions of the neural network model 360 and 370 start as copies of a common neural network configuration, e.g. that is initialized based on the first configuration data 380. Thus, just the parameters of the second version of the neural network model 370 change during training. The updated parameters is stored in the data storage device 372. An output from the first version of the neural network model 360 forms part of a loss function for the second version of the neural network model 370, such that during training the second version of the neural network model 370 “learns” to approximate the first version of the neural network model 360.
In [Col 19, lines 58-67], In [Col 20, lines 1-3] :
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570
- wherein updating, by the processing circuitry of the computing device, the learner unit by fitting, into the third fitted label set, the second statistical information further comprises further training the trained machine learning model with the at least one second residual and the first feature set;
In [Col 19, lines 58-67], In [Col 20, lines 1-3] :
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570. As per FIG. 4A or 4B, each of the versions output log its that are compared in a loss function to steer the gradient descent update of the parameters of the first version of the neural network model 562. Following training in this manner, the parameter data in the data storage device 560 is updated, allowing a revised or updated set of first configuration data 580 to be generated. This can then be communicated to a set of slave devices to repeat the process.
- third statistical information further defined by at least one third residual from fitting the third fitted label set.
In [Col 6, lines 47-55]:
In accordance with various aspects of the present invention, the receiving includes receiving second configuration data from a plurality of slave devices, and the updating includes: instantiating an ensemble of second versions of the neural network model as an ensemble of teacher models, and using the ensemble of teacher models to train the student model. The first version of the neural network model may thus be improved using private data from a large variety of heterogeneous devices.
The slave devices 722, 724, and 726 are communicatively coupled to the master device 710 via one or more communication networks 750. The master device 710 is configured to transmit first configuration data 780 for distribution to the plurality of slave devices 722, 724, and 726 via one or more communication networks 750.
In [Col 15, lines 15-19]:
Referring now to FIG. 2, a distributed training system 200 is shown in accordance with an embodiment of the invention. The distributed training system 200 may be seen as a variation of the first distributed training system 100 of FIG. 1 and similar features share similar reference numerals
In [Col 22, lines 48-63]:
a loss function may be a function of log it values across the ensemble of
instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.). Alternatively, the first version of the neural network model may be trained iteratively with each of the instantiated second versions being used as teachers in turn. In certain examples, the master device is able to evaluate and exclude slave configuration data from its teacher ensemble based on a predefined level of performance on a validation dataset, e.g. if an instantiated second version of the neural network model has an accuracy score below a threshold on the validation dataset, it may be excluded from the ensemble.
In [Col 17, lines 41-52]:
the first and second versions of the neural network model 360 and 370 start as copies of a common neural network configuration, e.g. that is initialized based on the first configuration data 380. Thus, just the parameters of the second version of the neural network model 370 change during training. The updated parameters is stored in the data storage device 372. An output from the first version of the neural network model 360 forms part of a loss function for the second version of the neural network model 370, such that during training the second version of the neural network model 370 “learns” to approximate the first version of the neural network model 360.
In [Col 19, lines 58-67], In [Col 20, lines 1-3] :
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570
(BRI: the plurality of fitted data is within the context of training plurality of slave devices. The plurality of statistical information relates to the plurality of slave devices that send the information)
- wherein sending, by the processing circuitry of the computing device, to the at least one module in the machine learning architecture,
In [Col 22, lines 49-53]:
a loss function may be a function of log it values across the ensemble of instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.)
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating the parameter data for the first version of the neural network model based on the second configuration data.
In [Col 22, lines 48-63]:
a loss function may be a function of log it values across the ensemble of
instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.).
In [Col 28, line 67], In [Col 298, lines 1-7]:
the receiving may include receiving second configuration data from a plurality of slave devices, e.g. as shown in FIG. 7 or FIGS. 8A to 8C. The updating at step 1120 may then include instantiating an ensemble of second versions of the neural network model as an ensemble of teacher models and using the ensemble of teacher models to train the student model. This is illustrated by the dotted lines in FIG. 5.
In [Col 17, lines 41-52]:
the first and second versions of the neural network model 360 and 370 start as copies of a common neural network configuration, e.g. that is initialized based on the first configuration data 380. Thus, just the parameters of the second version of the neural network model 370 change during training. The updated parameters is stored in the data storage device 372. An output from the first version of the neural network model 360 forms part of a loss function for the second version of the neural network model 370, such that during training the second version of the neural network model 370 “learns” to approximate the first version of the neural network model 360.
(BRI: the output from the first version of network for loss function to the second version is the residual and it implies that the first feature set is excluded)
- and further comprising sending, by the processing circuitry of the computing device, to the at least one module in the machine learning architecture,
In [Col 22, lines 49-53]:
a loss function may be a function of log it values across the ensemble of instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.)
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating the parameter data for the first version of the neural network model based on the second configuration data.
In [Col 22, lines 48-63]:
a loss function may be a function of log it values across the ensemble of
instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.).
In [Col 28, line 67], In [Col 298, lines 1-7]:
the receiving may include receiving second configuration data from a plurality of slave devices, e.g. as shown in FIG. 7 or FIGS. 8A to 8C. The updating at step 1120 may then include instantiating an ensemble of second versions of the neural network model as an ensemble of teacher models and using the ensemble of teacher models to train the student model. This is illustrated by the dotted lines in FIG. 5.
In [Col 17, lines 41-52]:
the first and second versions of the neural network model 360 and 370 start as copies of a common neural network configuration, e.g. that is initialized based on the first configuration data 380. Thus, just the parameters of the second version of the neural network model 370 change during training. The updated parameters is stored in the data storage device 372. An output from the first version of the neural network model 360 forms part of a loss function for the second version of the neural network model 370, such that during training the second version of the neural network model 370 “learns” to approximate the first version of the neural network model 360.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, and Li.
McMahan teaches providing user privacy.
Li teaches learning unit.
One of ordinary skill would have motivation to combine McMahan and Li that can improve the versions of neural network model without exchanging any private information (Li [Col 2, lines 54-66])
In regard to claim 9: (Original)
McMahan does not explicitly disclose:
- sending, by the processing circuitry of the computing device,
- the third statistical information further comprises determining a third particular residual based on a first third fitted label and at least one of the second particular residual of the at least one second residual, the first initial label of the initial label set, or the first observed data in the first feature set.
However, Li discloses:
- sending, by the processing circuitry of the computing device,
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating the parameter data for the first version of the neural network model based on the second configuration data.
- the third statistical information further comprises determining a third particular residual based on a first third fitted label and at least one of the second particular residual of the at least one second residual, the first initial label of the initial label set, or the first observed data in the first feature set.
In [Col 17, lines 41-52]:
the first and second versions of the neural network model 360 and 370 start as copies of a common neural network configuration, e.g. that is initialized based on the first configuration data 380. Thus, just the parameters of the second version of the neural network model 370 change during training. The updated parameters is stored in the data storage device 372. An output from the first version of the neural network model 360 forms part of a loss function for the second version of the neural network model 370, such that during training the second version of the neural network model 370 “learns” to approximate the first version of the neural network model 360.
In [Col 22, lines 49-53]:
a loss function may be a function of log it values across the ensemble of instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, and Li.
McMahan teaches providing user privacy.
Li teaches learning unit.
One of ordinary skill would have motivation to combine McMahan and Li that can improve the versions of neural network model without exchanging any private information (Li [Col 2, lines 54-66])
In regard to claim 10: (Currently Amended)
McMahan discloses:
- A computing device comprising: processing circuitry coupled to memory and configured to:
In [0004]:
One example aspect of the present disclosure is directed to a computer-implemented method of updating a global model based on unevenly distributed data. The method includes receiving, by one or more computing devices, one or more local updates from a plurality of user devices. Each local update is determined by the respective user device based at least in part on one or more data examples stored on the respective user device. The one or more data examples stored on the plurality of user devices are distributed on an uneven basis, such that no user device includes a representative sample of an overall distribution of data examples. The method further includes aggregating, by the one or more computing devices
in [0054]:
The technology discussed herein makes reference to servers, databases, software applications, and other computer-based systems, as well as actions taken and information sent to and from such systems.
- wherein each of the at least one module executes on at least one remote computing device, and wherein the computing device maintains privacy in sending the first statistical information
In [0015]:
More particularly, in some embodiments, a set of input-output data can be used to describe a global objective via a loss function. Such functions can be, for instance, a convex or non-convex function, such as a linear regression function, logistic regression function. A local objective (F.sub.k) can also be defined using data stored on a computing device. For instance, the global objective can be defined as:
(BRI: It is known in that a linear regression is considered a statistical model; it is a method used to estimate the relationship between a dependent variable and one or more independent variables by assuming a relationship between them, allowing for predictions based on this model and statistical analysis of the data)
In [0030]:
FIG. 1 depicts an example system 300 for training one or more global machine learning models 306 using training data 308 stored locally on a plurality of user devices 302. System 300 can further include a server device 304. Server 304 can be configured to access machine learning model 306, and to provide model 306 to a plurality of user devices 302. Model 306 can be, for instance, a linear regression model,
In [0030]:
in some implementations, sever 304 can be configured to communicate with user devices 302 over one or more networks, such as network 240 of FIG. 3.
In [0052]:
The client device 230 can also include a network interface used to communicate with one or more remote computing devices (e.g. server 210) over the network 240.
in [0016]:
According to particular implementations of the present disclosure, the global objective can be solved by aggregating a plurality of local updates provided by a plurality of remote computing devices. Each remote computing device can, for instance, be a user device, such as a laptop computing device, desktop computing device, smartphone, tablet, wearable computing device, or other computing device. The local updates can be determined based at least in part on the respective local objectives.
in [0054]:
Databases and applications may be implemented on a single system or distributed across multiple systems.
in [0014]:
the local update may be determined using one or more gradient descent techniques
in [0014]:
the local update does not include the training data used to determine the local update. In this manner, the size of the local update can be independent of the training data used to determine the local update, thereby reducing bandwidth requirements and maintaining user privacy. In particular, a global model can be updated based at least in part on the received local updates. By only providing the local update (and not the training data) to the server, the global model update can be determined using reduced bandwidth requirements, and without compromising the security of potentially privacy sensitive data stored on the user devices
(BRI: It is known in the art that data stored on a user device can be considered a feature set, especially when used for machine learning or data analysis purposes, as it can be used to train models or make predictions)
McMahan does not explicitly disclose:
- create a learner unit by fitting, into a first fitted label set,
- an initial label set using at least one first learning technique, a first machine learning model, and a first feature set;
- first statistical information defined by at least one first residual from fitting the first fitted label set, the first statistical information that includes at least one first residual from fitting the first fitted label set that excludes both the first feature set and first model information of the first model, wherein module is operative to fit, into second fitted label set ,wherein module is operative to fit, into at least one second fitted label set
- and update the learner unit by fitting, into a third fitted label set, the second statistical information using the at least one first learning technique and the first machine learning model.
However, Li discloses:
- create a learner unit by fitting, into a first fitted label set,
In [Col 4, lines 41-56]:
In accordance with various aspects of the present invention, the first configuration data includes hyperparameters for the neural network model and parameters for the first version of the neural network model. The hyperparameters may include one or more of: an architecture definition for the neural network model; a number of nodes for one or more layers in the neural network model; a set of node definitions indicating one or more of a node type and a node connectivity; a set of activation function definitions; and one or more cost function definitions. The parameters may include one or more of: weight values for one or more connections between nodes of the neural network model; weight values for one or more inputs to the neural network model; weight values for one or more recurrent paths in the neural network model; and bias values for one or more nodes of the neural network model.
(BRI: A learner unit is basic computational element that performs a calculation on its inputs to produce an output and passing the result through an activation function, enabling the network to learn complex patterns in data during the training process)
In [ Col 9, lines 17-23]:
“neural network model” is used to refer to an artificial neural network that is configured to perform a particular data processing task. For example, in the case that a neural network model includes an acoustic model, the task may be to output phoneme or grapheme data (e.g. predictions of phonemes or graphemes) based on input audio data.
In [ Col 9, lines 26-32]:
In certain cases, a neural network model may be a model that is configured to provide a particular mapping between defined input data and defined output data. The input data may represent one modality and output data may represent another modality. The neural network model may be considered a function approximator that is trained on a set of data.
(BRI: A "fitted label set" is a concept most relevant to machine learning, where it refers to the mapping of numerical labels to categorical data after a model has been "fitted" or trained. A modality can act as a type of "label" or category to describe a type of data)
In [Col 19, lines 58-62]:
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570.
- an initial label set using at least one first learning technique, a first machine learning model, and a first feature set;
In [Col 12, lines 2-6]:
a version of the neural network model may be instantiated by implementing a class or class-like definition of the neural network model using initialization data. In this case, the initialization data includes the first configuration data 180.
- send to at least one module in a machine learning architecture,
In [Col 22, lines 49-53]:
a loss function may be a function of log it values across the ensemble of instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.)
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating the parameter data for the first version of the neural network model based on the second configuration data.
- receive, from the at least one module,
In [Col 22, lines 49-53]:
a loss function may be a function of log it values across the ensemble of instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.)
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating the parameter data for the first version of the neural network model based on the second configuration data.
- and update the learner unit by fitting, into a third fitted label set, the second statistical information using the at least one first learning technique and the first machine learning model.
In [Col 19, lines 58-67], In [Col 20, lines 1-3] :
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570. As per FIG. 4A or 4B, each of the versions output log its that are compared in a loss function to steer the gradient descent update of the parameters of the first version of the neural network model 562. Following training in this manner, the parameter data in the data storage device 560 is updated, allowing a revised or updated set of first configuration data 580 to be generated. This can then be communicated to a set of slave devices to repeat the process.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, and Li.
McMahan teaches providing user privacy.
Li teaches data set of fit, label and plurality of models and a learning unit.
One of ordinary skill would have motivation to combine McMahan and Li that can improve the versions of neural network model without exchanging any private information (Li [Col 2, lines 54-66])
However, Zadeh discloses:
- first statistical information that includes at least one first residual from fitting the first fitted label set that excludes both the first feature set, and first model information of the first machine model, from disclosure to the at least one module, wherein module is operative to fit, into at least one second fitted label set
[3162]:
the learning process of the learning machine uses a statistical approach to hypothesis parameters by keeping track of those values over time during fitting the data fitting, with more weight given to those parameters occurring more often.
(BRI: by tracking parameter values over time and weighting them represents aggregating and reinforcing statistical information from multiple data points, and represents of plurality of statistical information. In this context, the statistical information is a first statistical information)
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
[0565]:
The membership function of A, μ.sub.A, may be elicited by asking a succession of questions of the form: To what degree does the number, a, fit your perception of A? Example: To what degree does 50 minutes fit your perception of about 45 minutes? The same applies to B. The fuzzy set, A, may be interpreted as the possibility distribution of X. The concept of a Z-number may be generalized in various ways. In particular, X may be assumed to take values in R.sup.n, in which case A is a Cartesian product of fuzzy numbers.
[1638]:
An approach to approximate or render the overlap (15) between the category sets, such as C.sub.X,n, may use α-cuts to present each crisp α-cuts of predetermined category set as a set of points in (m,σ) space. These sets of points may be modeled efficiently, e.g., based on graphical models, optimized for fast transformation and intersection operations. For example, the models that use peripheral description for the α-cuts allow robust and efficient determination of intersection and avoid the need to transform all the points within the set individually, in order to reduce the computation involved in (13).
(BRI: Elements and their degree of membership and within the context of A and B as membership. As there is overlapping between categories sets (fuzzy sets), it does represent plurality of fitted label sets because an element can have multiple non-zero membership degrees, each indicating the extent to which it fits into a different categories)
[2129]:
For the aggregation method (also called ensemble learning, or boosting, or mixture of experts), we have a learning which tries to replicate the function independently (not jointly), and then combine and put them together later,
[2129]:
For the aggregation method, for regression or real number cases, we take an average or weighted average,
[2129]:
For the aggregation method, we have 2 types: (a) After-the-fact situation (where we already have the solutions, and then we combine them), and (b) Before-the-fact situation (where we get solutions, with the view or intention or assumption to blend or combine them together later). For the aggregation method, as one example, we have the Boosting method, where we enforce the decorrelation (not by chance), e.g. by building one hypothesis at a time, for a good mixture, sequentially.
[2092] :
In one embodiment, for aggregating the correlated fuzzy sets, for the Min-Max aggregation method, we can get the final membership value, μ.sub.final, based on the individual membership values, μ.sub.i and μ.sub.2, as (where index i runs from 0 to n):
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership, the set is the fitted labeled set)
[1723]:
In one embodiment, the added units and the previous units are used to make association and/or correlation with labeled samples, e.g., during the supervised training.
[1723]:
In one embodiment, the labels are continuous valued (or multi-valued), e.g., having values in range [0, 1], to indicate the degree in which the document is classified by a label (or the membership function of the document in the label's class).
[0113]:
For one embodiment: Decisions are based on information. To be useful, information must be reliable. Basically, the concept of a Z-number relates to the issue of reliability of information. A Z-number, Z, has two components, Z=(A,B). The first component, A, is a restriction (constraint) on the values which a real-valued uncertain variable, X, is allowed to take. The second component, B, is a measure of reliability (certainty) of the first component.
[0564]:
For purposes of computation, when A and B are described in a natural language, the meaning of A and B is precisiated (graduated) through association with membership functions, μ.sub.A and μ.sub.B, respectively, FIG. 1.
[0593]:
A basic fuzzy if-then rule may be expressed as: if X is A then Y is B, where A and B are fuzzy numbers. The meaning of such a rule is defined as:
PNG
media_image2.png
17
202
media_image2.png
Greyscale
[1678] :
In one embodiment, the learning is achieved through simulation using a data (and label) sample generation based on one or more models. In one embodiment, a network trained based on model(s) is used to recognize and classify actual data which may not have been seen before. In one embodiment, the system is trained to infer the potential model(s) itself by recognizing the (e.g., observed) data conforming to a particular model and its associated labels/parameters.
[1757]:
The features of both images (and in particular the differences between their features) are correlated using a correlator/analyzer module (e.g., having unit/neurons) with the label/feature difference identified in the label layer (e.g., L.sub.1). In one embodiment, the L layer represents the labels indicated the feature differences between the images (denoted by ΔL). In one embodiment, more than one label is selected by the controller (indicating the differences between the features of the images selected from the database for training the correlation).
[2431]:
In one embodiment, the system uses statistics and patterns extracted from images
[2788]:
a script is executed that iterate through the document object model or the window object to identify elements associated with images
[1723]:
In one embodiment, feature detection system is used to train document classification based on learned (e.g., unsupervised) features corresponding to documents based on terms contained in the document (such as statistics of several hundred or several thousand common words)
[2521]:
In one embodiment, as for example depicted in FIG. 228, based on an identifier or a URL, a movie and its associated data (e.g., metadata, title, description, owner/uploader, channel, comments, likes, and statistics) are extracted from a repository. In one embodiment, the movie and its associated data are analyzed, e.g., by a video/audio analyzer and keyword/concept extraction/analyzer, to provide/generate features/annotation and metadata
[1757]:
In one embodiment, the weights related to the units in the correlator/analyzer are trained to detect the feature differences by a stochastic or batch learning algorithm.
[1734]:
features of an object (e.g., pose including rotation) is determined, and based on such features, features of sub-objects of other objects depicted in an image are extracted by preprocessing (e.g., mapping) a portion of an image into a segmented layout with variable resolution. Then, the mapped image (or portion thereof) is provided to a classifier or feature recognition system to determine the features from the mapped image
[1351]:
In one embodiment, an optimum statistical classifier is used. In one embodiment, a Bayes classifier is used
[1351]:
one embodiment, a perceptron for 2-pattern classes is used. In one embodiment, the least mean square (LMS) delta rule for training perceptrons is used, to minimize the error between the actual response and the desired response (for the training purposes). FIG. 115 is an example of a system described above.
[1352] :
In one embodiment, a multi-layer feed-forward neural network is used. In one embodiment, the training is done by back propagation, using the total squared error between the actual responses and desired responses for the nodes in the output layer.
[2301]:
In one embodiment, we segment the data of any type, including video, sound, and multimedia, based on sudden change in the sequence (or big delta or difference),
[2300]:
In one embodiment, we use Bayesian model, for both sides of the potential boundary between segments, with 2 different model parameters, to fit the 2 sides better, to examine the potential boundary for segmentation, e.g. for speech.
(BRI: in the context of a neural network or any predictive model, the difference between the actual output (response) and the desired output (target) for the nodes in the output layer is indeed the residual. From a statistical information perspective, these residuals are not just raw differences and they are statistical information in the sense that they quantify the discrepancy between the model’s predictions and the true data. The plurality of residuals are within the context of residual associated to each node and the statistical information associated to each residual)
[1768] :
In one embodiment, meta data such as the GPS data (or for example other accompanying metadata captured with images taken from mobile devices such as smart phones) are used as labels (e.g., continuous valued).
[1586]:
The communication between different units, devices, or modules are done by wire, cable, fiber optics, wirelessly, WiFi, Bluetooth, through network, Internet, copper interconnect, antenna, satellite dish, or the like.
1759]:
In one embodiment, one or more pose detection modules (e.g., based on edge detection or color region/shape) are used to determine the pose of a face within an image/data
[1810]:
FIG. 127 shows a system for context determination, with language input device, which feeds dissecting and parsing modules to get the components or parts of the sentence, which feeds the analyzing module (which e.g. may include memory units and processor units or CPU or computing module), which is connected to the context determination module, which is connected to the default analyzer module and multiple other context analyzer module
PNG
media_image3.png
537
513
media_image3.png
Greyscale
[2422]:
In one embodiment, we have private or public or semi-public or semi-private (or the like) settings for our sharing or displaying or reviewing or tagging or annotating or accessing or searching or browsing of images or objects or videos, for user, friends, family, co-worker, boss, employee, contractor, senior management, public, social network, college, school, classmate, roommate, household, shared device, shared account, friend-of-friend, friend-of-friend-of-friend, and so on, or the like. In one embodiment, we have government excluding list database, for specified individuals, to exclude for the rules, for the above functions, for privacy settings. In one embodiment, we have the intersection of the privacy settings of multiple users or contributors
[2789]:
In one embodiment, the script determines whether to identify the images based on the domain name (e.g., as positive filter to include or negative filter to exclude).
[2830]:
publisher's restriction/filters to exclude certain or certain types of merchants and merchant's restriction/filters to exclude certain or certain types of publishers.
[1917]:
In one embodiment, for conditional relationships, or multiple choices, we can continue, until we get to a dead end or conflict, and then, backtrack to eliminate or adjust one or more choices, on the chain going backward, to correct or adjust some assumptions, choices, or conditions, on the way.
[2355]:
we can input these conditions or rules into our rule engine, or use it for prediction, control system, forecasting (economy, elections, and other events), social behavioral analysis, consumer behavioral analysis, predicting revolutions or unrest, detecting frauds, detecting unusual behaviors, detecting unusual patterns, finding liars or contradictions
- second statistical information first statistical information that includes at least one second residual from fitting the second fitted label set that excludes both the second feature set, and second model information of the second machine model, and that is based on the first statistical information, from disclosure to the at least one module, and that is based on the first statistical information;
[3162]:
the learning process of the learning machine uses a statistical approach to hypothesis parameters by keeping track of those values over time during fitting the data fitting, with more weight given to those parameters occurring more often.
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
[0565]:
The membership function of A, μ.sub.A, may be elicited by asking a succession of questions of the form: To what degree does the number, a, fit your perception of A? Example: To what degree does 50 minutes fit your perception of about 45 minutes? The same applies to B. The fuzzy set, A, may be interpreted as the possibility distribution of X. The concept of a Z-number may be generalized in various ways. In particular, X may be assumed to take values in R.sup.n, in which case A is a Cartesian product of fuzzy numbers.
[1638]:
An approach to approximate or render the overlap (15) between the category sets, such as C.sub.X,n, may use α-cuts to present each crisp α-cuts of predetermined category set as a set of points in (m,σ) space. These sets of points may be modeled efficiently, e.g., based on graphical models, optimized for fast transformation and intersection operations. For example, the models that use peripheral description for the α-cuts allow robust and efficient determination of intersection and avoid the need to transform all the points within the set individually, in order to reduce the computation involved in (13).
(BRI: Elements and their degree of membership and within the context of A and B as membership. As there is overlapping between categories sets (fuzzy sets), it does represent plurality of fitted label sets because an element can have multiple non-zero membership degrees, each indicating the extent to which it fits into a different categories)
[2129]:
For the aggregation method (also called ensemble learning, or boosting, or mixture of experts), we have a learning which tries to replicate the function independently (not jointly), and then combine and put them together later,
[2129]:
For the aggregation method, for regression or real number cases, we take an average or weighted average,
[2129]:
For the aggregation method, we have 2 types: (a) After-the-fact situation (where we already have the solutions, and then we combine them), and (b) Before-the-fact situation (where we get solutions, with the view or intention or assumption to blend or combine them together later). For the aggregation method, as one example, we have the Boosting method, where we enforce the decorrelation (not by chance), e.g. by building one hypothesis at a time, for a good mixture, sequentially.
[2092] :
In one embodiment, for aggregating the correlated fuzzy sets, for the Min-Max aggregation method, we can get the final membership value, μ.sub.final, based on the individual membership values, μ.sub.i and μ.sub.2, as (where index i runs from 0 to n):
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership, the set is the fitted labeled set)
[1723]:
In one embodiment, the added units and the previous units are used to make association and/or correlation with labeled samples, e.g., during the supervised training.
[1723]:
In one embodiment, the labels are continuous valued (or multi-valued), e.g., having values in range [0, 1], to indicate the degree in which the document is classified by a label (or the membership function of the document in the label's class).
[0113]:
For one embodiment: Decisions are based on information. To be useful, information must be reliable. Basically, the concept of a Z-number relates to the issue of reliability of information. A Z-number, Z, has two components, Z=(A,B). The first component, A, is a restriction (constraint) on the values which a real-valued uncertain variable, X, is allowed to take. The second component, B, is a measure of reliability (certainty) of the first component.
[0564]:
For purposes of computation, when A and B are described in a natural language, the meaning of A and B is precisiated (graduated) through association with membership functions, μ.sub.A and μ.sub.B, respectively, FIG. 1.
[0593]:
A basic fuzzy if-then rule may be expressed as: if X is A then Y is B, where A and B are fuzzy numbers. The meaning of such a rule is defined as:
PNG
media_image2.png
17
202
media_image2.png
Greyscale
[1678] :
In one embodiment, the learning is achieved through simulation using a data (and label) sample generation based on one or more models. In one embodiment, a network trained based on model(s) is used to recognize and classify actual data which may not have been seen before. In one embodiment, the system is trained to infer the potential model(s) itself by recognizing the (e.g., observed) data conforming to a particular model and its associated labels/parameters.
[1757]:
The features of both images (and in particular the differences between their features) are correlated using a correlator/analyzer module (e.g., having unit/neurons) with the label/feature difference identified in the label layer (e.g., L.sub.1). In one embodiment, the L layer represents the labels indicated the feature differences between the images (denoted by ΔL). In one embodiment, more than one label is selected by the controller (indicating the differences between the features of the images selected from the database for training the correlation).
[2431]:
In one embodiment, the system uses statistics and patterns extracted from images
[2788]:
a script is executed that iterate through the document object model or the window object to identify elements associated with images
[1723]:
In one embodiment, feature detection system is used to train document classification based on learned (e.g., unsupervised) features corresponding to documents based on terms contained in the document (such as statistics of several hundred or several thousand common words)
[2521]:
In one embodiment, as for example depicted in FIG. 228, based on an identifier or a URL, a movie and its associated data (e.g., metadata, title, description, owner/uploader, channel, comments, likes, and statistics) are extracted from a repository. In one embodiment, the movie and its associated data are analyzed, e.g., by a video/audio analyzer and keyword/concept extraction/analyzer, to provide/generate features/annotation and metadata
[1757]:
In one embodiment, the weights related to the units in the correlator/analyzer are trained to detect the feature differences by a stochastic or batch learning algorithm.
[1734]:
features of an object (e.g., pose including rotation) is determined, and based on such features, features of sub-objects of other objects depicted in an image are extracted by preprocessing (e.g., mapping) a portion of an image into a segmented layout with variable resolution. Then, the mapped image (or portion thereof) is provided to a classifier or feature recognition system to determine the features from the mapped image
[1351]:
In one embodiment, an optimum statistical classifier is used. In one embodiment, a Bayes classifier is used
[1351]:
one embodiment, a perceptron for 2-pattern classes is used. In one embodiment, the least mean square (LMS) delta rule for training perceptrons is used, to minimize the error between the actual response and the desired response (for the training purposes). FIG. 115 is an example of a system described above.
[1352] :
In one embodiment, a multi-layer feed-forward neural network is used. In one embodiment, the training is done by back propagation, using the total squared error between the actual responses and desired responses for the nodes in the output layer.
[2301]:
In one embodiment, we segment the data of any type, including video, sound, and multimedia, based on sudden change in the sequence (or big delta or difference),
[2300]:
In one embodiment, we use Bayesian model, for both sides of the potential boundary between segments, with 2 different model parameters, to fit the 2 sides better, to examine the potential boundary for segmentation, e.g. for speech.
(BRI: in the context of a neural network or any predictive model, the difference between the actual output (response) and the desired output (target) for the nodes in the output layer is indeed the residual. From a statistical information perspective, these residuals are not just raw differences and they are statistical information in the sense that they quantify the discrepancy between the model’s predictions and the true data. The plurality of residuals are within the context of residual associated to each node and the statistical information associated to each residual)
[1734]:
in one embodiment, based on the perspective/skew/projection of the frame (or other indicators), the image or a portion of image is mapped to a segmented layout for input to a network for further feature detection or classification.
[1765]:
to detect the object/class of object based on a thumbnail (e.g., via preprocessing) of the data/image. In one embodiment, the training for a high level feature detection focuses on the structure of the neurons or units used in a classifier/feature detector. In one embodiment, the resulting feature units at top layer are limited to few features, while the training is used with data/images that may include thumbnail and high resolution data/images, including those with and without the targeted features.
[1768] :
In one embodiment, meta data such as the GPS data (or for example other accompanying metadata captured with images taken from mobile devices such as smart phones) are used as labels (e.g., continuous valued).
[1586]:
The communication between different units, devices, or modules are done by wire, cable, fiber optics, wirelessly, WiFi, Bluetooth, through network, Internet, copper interconnect, antenna, satellite dish, or the like.
1759]:
In one embodiment, one or more pose detection modules (e.g., based on edge detection or color region/shape) are used to determine the pose of a face within an image/data
[1810]:
FIG. 127 shows a system for context determination, with language input device, which feeds dissecting and parsing modules to get the components or parts of the sentence, which feeds the analyzing module (which e.g. may include memory units and processor units or CPU or computing module), which is connected to the context determination module, which is connected to the default analyzer module and multiple other context analyzer module
PNG
media_image3.png
537
513
media_image3.png
Greyscale
[2422]:
In one embodiment, we have private or public or semi-public or semi-private (or the like) settings for our sharing or displaying or reviewing or tagging or annotating or accessing or searching or browsing of images or objects or videos, for user, friends, family, co-worker, boss, employee, contractor, senior management, public, social network, college, school, classmate, roommate, household, shared device, shared account, friend-of-friend, friend-of-friend-of-friend, and so on, or the like. In one embodiment, we have government excluding list database, for specified individuals, to exclude for the rules, for the above functions, for privacy settings. In one embodiment, we have the intersection of the privacy settings of multiple users or contributors
[2789]:
In one embodiment, the script determines whether to identify the images based on the domain name (e.g., as positive filter to include or negative filter to exclude).
[2830]:
publisher's restriction/filters to exclude certain or certain types of merchants and merchant's restriction/filters to exclude certain or certain types of publishers.
[1917]:
In one embodiment, for conditional relationships, or multiple choices, we can continue, until we get to a dead end or conflict, and then, backtrack to eliminate or adjust one or more choices, on the chain going backward, to correct or adjust some assumptions, choices, or conditions, on the way.
[2355]:
we can input these conditions or rules into our rule engine, or use it for prediction, control system, forecasting (economy, elections, and other events), social behavioral analysis, consumer behavioral analysis, predicting revolutions or unrest, detecting frauds, detecting unusual behaviors, detecting unusual patterns, finding liars or contradictions
- update the learner unit by fitting, into a third fitted label set, the second statistical information using the at least one first learning technique and the first machine learning model.
[1340]:
the fuzzy classifier module or device classifies or separates different pictures into clusters or groups in N-dimensional feature space.
[1531]:
a method for fuzzy logic control, in which an input module receives a precisiated proposition associated with a protoform. A fuzzy logic inference engine evaluates a first fuzzy logic rule from the fuzzy logic rule repository. The fuzzy logic inference engine is in or loaded on or executed on or implemented in a computing device, which comprises one or more of following: computer, processor device
[0565]:
The membership function of A, μ.sub.A, may be elicited by asking a succession of questions of the form: To what degree does the number, a, fit your perception of A? Example: To what degree does 50 minutes fit your perception of about 45 minutes? The same applies to B. The fuzzy set, A, may be interpreted as the possibility distribution of X. The concept of a Z-number may be generalized in various ways. In particular, X may be assumed to take values in R.sup.n, in which case A is a Cartesian product of fuzzy numbers. Simple examples of Z-valuations are:
(BRI: When fitting a model to data, the goal is to find parameters that best match observed values. With fuzzy data, the “fit” is adjusted to account for uncertainty)
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership, the set is the fitted labeled set)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
In regard to claim 11: (Original)
McMahan does not explicitly disclose:
- wherein the processing circuitry is further configured to:
- send to the at least one module in the machine learning architecture, third statistical information defined by at least one third residual from fitting the third fitted label set using the at least one first learning technique and the first machine learning model, wherein the at least module is operative to fit, into a fourth fitted label set, the third statistical information using the at least one second learning technique, the at least one second machine learning model, and the at least one second feature set.
However, Li discloses:
- the processing circuitry is further configured to: send to the at least one module in the machine learning architecture, wherein the at least module is operative to fit
In [Col 22, lines 49-53]:
a loss function may be a function of log it values across the ensemble of instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.)
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating
In [Col 22, lines 48-63]: the parameter data for the first version of the neural network model based on the second configuration data.
a loss function may be a function of log it values across the ensemble of
instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.).
In [Col 28, line 67], In [Col 298, lines 1-7]:
the receiving may include receiving second configuration data from a plurality of slave devices, e.g. as shown in FIG. 7 or FIGS. 8A to 8C. The updating at step 1120 may then include instantiating an ensemble of second versions of the neural network model as an ensemble of teacher models and using the ensemble of teacher models to train the student model. This is illustrated by the dotted lines in FIG. 5.
In [Col 17, lines 41-52]:
the first and second versions of the neural network model 360 and 370 start as copies of a common neural network configuration, e.g. that is initialized based on the first configuration data 380. Thus, just the parameters of the second version of the neural network model 370 change during training. The updated parameters is stored in the data storage device 372. An output from the first version of the neural network model 360 forms part of a loss function for the second version of the neural network model 370, such that during training the second version of the neural network model 370 “learns” to approximate the first version of the neural network model 360.
(BRI: the output from the first version of network for loss function to the second version is the residual and it implies that the first feature set is excluded)
McMahan and Li do not explicitly disclose:
- third statistical information defined by at least one third residual from fitting the third fitted label set using the at least one first learning technique and the first machine learning model
- wherein the at least module is operative to fit, into a fourth fitted label set, the third statistical information using the at least one second learning technique, the at least one second machine learning model, and the at least one second feature set.
However, Zadeh discloses:
- third statistical information defined by at least one third residual from fitting the third fitted label set using the at least one first learning technique and the first machine learning model
[3162]:
the learning process of the learning machine uses a statistical approach to hypothesis parameters by keeping track of those values over time during fitting the data fitting, with more weight given to those parameters occurring more often.
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
[0565]:
The membership function of A, μ.sub.A, may be elicited by asking a succession of questions of the form: To what degree does the number, a, fit your perception of A? Example: To what degree does 50 minutes fit your perception of about 45 minutes? The same applies to B. The fuzzy set, A, may be interpreted as the possibility distribution of X. The concept of a Z-number may be generalized in various ways. In particular, X may be assumed to take values in R.sup.n, in which case A is a Cartesian product of fuzzy numbers.
[1638] :
An approach to approximate or render the overlap (15) between the category sets, such as C.sub.X,n, may use α-cuts to present each crisp α-cuts of predetermined category set as a set of points in (m,σ) space. These sets of points may be modeled efficiently, e.g., based on graphical models, optimized for fast transformation and intersection operations. For example, the models that use peripheral description for the α-cuts allow robust and efficient determination of intersection and avoid the need to transform all the points within the set individually, in order to reduce the computation involved in (13).
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership. As there is overlapping between categories sets (fuzzy sets), it does represent plurality of fitted label sets because an element can have multiple non-zero membership degrees, each indicating the extent to which it fits into a different categories)
[2129]:
For the aggregation method (also called ensemble learning, or boosting, or mixture of experts), we have a learning which tries to replicate the function independently (not jointly), and then combine and put them together later,
[2129]:
For the aggregation method, for regression or real number cases, we take an average or weighted average,
[2129]:
For the aggregation method, we have 2 types: (a) After-the-fact situation (where we already have the solutions, and then we combine them), and (b) Before-the-fact situation (where we get solutions, with the view or intention or assumption to blend or combine them together later). For the aggregation method, as one example, we have the Boosting method, where we enforce the decorrelation (not by chance), e.g. by building one hypothesis at a time, for a good mixture, sequentially.
[2092] :
In one embodiment, for aggregating the correlated fuzzy sets, for the Min-Max aggregation method, we can get the final membership value, μ.sub.final, based on the individual membership values, μ.sub.i and μ.sub.2, as (where index i runs from 0 to n):
[1723]:
In one embodiment, the added units and the previous units are used to make association and/or correlation with labeled samples, e.g., during the supervised training.
[1723]:
In one embodiment, the labels are continuous valued (or multi-valued), e.g., having values in range [0, 1], to indicate the degree in which the document is classified by a label (or the membership function of the document in the label's class).
[0113]:
For one embodiment: Decisions are based on information. To be useful, information must be reliable. Basically, the concept of a Z-number relates to the issue of reliability of information. A Z-number, Z, has two components, Z=(A,B). The first component, A, is a restriction (constraint) on the values which a real-valued uncertain variable, X, is allowed to take. The second component, B, is a measure of reliability (certainty) of the first component.
[0564]:
For purposes of computation, when A and B are described in a natural language, the meaning of A and B is precisiated (graduated) through association with membership functions, μ.sub.A and μ.sub.B, respectively, FIG. 1.
[0593]:
A basic fuzzy if-then rule may be expressed as: if X is A then Y is B, where A and B are fuzzy numbers. The meaning of such a rule is defined as:
PNG
media_image2.png
17
202
media_image2.png
Greyscale
[1678] :
In one embodiment, the learning is achieved through simulation using a data (and label) sample generation based on one or more models. In one embodiment, a network trained based on model(s) is used to recognize and classify actual data which may not have been seen before. In one embodiment, the system is trained to infer the potential model(s) itself by recognizing the (e.g., observed) data conforming to a particular model and its associated labels/parameters.
[1757]:
The features of both images (and in particular the differences between their features) are correlated using a correlator/analyzer module (e.g., having unit/neurons) with the label/feature difference identified in the label layer (e.g., L.sub.1). In one embodiment, the L layer represents the labels indicated the feature differences between the images (denoted by ΔL). In one embodiment, more than one label is selected by the controller (indicating the differences between the features of the images selected from the database for training the correlation).
[2431]:
In one embodiment, the system uses statistics and patterns extracted from images
[2788]:
a script is executed that iterate through the document object model or the window object to identify elements associated with images
[1723]:
In one embodiment, feature detection system is used to train document classification based on learned (e.g., unsupervised) features corresponding to documents based on terms contained in the document (such as statistics of several hundred or several thousand common words)
[2521]:
In one embodiment, as for example depicted in FIG. 228, based on an identifier or a URL, a movie and its associated data (e.g., metadata, title, description, owner/uploader, channel, comments, likes, and statistics) are extracted from a repository. In one embodiment, the movie and its associated data are analyzed, e.g., by a video/audio analyzer and keyword/concept extraction/analyzer, to provide/generate features/annotation and metadata
[1757]:
In one embodiment, the weights related to the units in the correlator/analyzer are trained to detect the feature differences by a stochastic or batch learning algorithm.
[1734]:
features of an object (e.g., pose including rotation) is determined, and based on such features, features of sub-objects of other objects depicted in an image are extracted by preprocessing (e.g., mapping) a portion of an image into a segmented layout with variable resolution. Then, the mapped image (or portion thereof) is provided to a classifier or feature recognition system to determine the features from the mapped image
[1351]:
In one embodiment, an optimum statistical classifier is used. In one embodiment, a Bayes classifier is used
[1351]:
one embodiment, a perceptron for 2-pattern classes is used. In one embodiment, the least mean square (LMS) delta rule for training perceptrons is used, to minimize the error between the actual response and the desired response (for the training purposes). FIG. 115 is an example of a system described above.
[1352] :
In one embodiment, a multi-layer feed-forward neural network is used. In one embodiment, the training is done by back propagation, using the total squared error between the actual responses and desired responses for the nodes in the output layer.
[2301]:
In one embodiment, we segment the data of any type, including video, sound, and multimedia, based on sudden change in the sequence (or big delta or difference),
[2300]:
In one embodiment, we use Bayesian model, for both sides of the potential boundary between segments, with 2 different model parameters, to fit the 2 sides better, to examine the potential boundary for segmentation, e.g. for speech.
(BRI: The plurality of residuals are within the context of residual associated to each node and the statistical information associated to each residual. In this context, the residual is the third residual associated to third statistical information)
[1765]:
to detect the object/class of object based on a thumbnail (e.g., via preprocessing) of the data/image. In one embodiment, the training for a high level feature detection focuses on the structure of the neurons or units used in a classifier/feature detector. In one embodiment, the resulting feature units at top layer are limited to few features, while the training is used with data/images that may include thumbnail and high resolution data/images, including those with and without the targeted features.
[1768] :
In one embodiment, meta data such as the GPS data (or for example other accompanying metadata captured with images taken from mobile devices such as smart phones) are used as labels (e.g., continuous valued).
[1586]:
The communication between different units, devices, or modules are done by wire, cable, fiber optics, wirelessly, WiFi, Bluetooth, through network, Internet, copper interconnect, antenna, satellite dish, or the like.
1759]:
In one embodiment, one or more pose detection modules (e.g., based on edge detection or color region/shape) are used to determine the pose of a face within an image/data
[1810]:
FIG. 127 shows a system for context determination, with language input device, which feeds dissecting and parsing modules to get the components or parts of the sentence, which feeds the analyzing module (which e.g. may include memory units and processor units or CPU or computing module), which is connected to the context determination module, which is connected to the default analyzer module and multiple other context analyzer module
PNG
media_image3.png
537
513
media_image3.png
Greyscale
[2422]:
In one embodiment, we have private or public or semi-public or semi-private (or the like) settings for our sharing or displaying or reviewing or tagging or annotating or accessing or searching or browsing of images or objects or videos, for user, friends, family, co-worker, boss, employee, contractor, senior management, public, social network, college, school, classmate, roommate, household, shared device, shared account, friend-of-friend, friend-of-friend-of-friend, and so on, or the like. In one embodiment, we have government excluding list database, for specified individuals, to exclude for the rules, for the above functions, for privacy settings. In one embodiment, we have the intersection of the privacy settings of multiple users or contributors
[2789]:
In one embodiment, the script determines whether to identify the images based on the domain name (e.g., as positive filter to include or negative filter to exclude).
[2830]:
publisher's restriction/filters to exclude certain or certain types of merchants and merchant's restriction/filters to exclude certain or certain types of publishers.
[1917]:
In one embodiment, for conditional relationships, or multiple choices, we can continue, until we get to a dead end or conflict, and then, backtrack to eliminate or adjust one or more choices, on the chain going backward, to correct or adjust some assumptions, choices, or conditions, on the way.
[2355]:
we can input these conditions or rules into our rule engine, or use it for prediction, control system, forecasting (economy, elections, and other events), social behavioral analysis, consumer behavioral analysis, predicting revolutions or unrest, detecting frauds, detecting unusual behaviors, detecting unusual patterns, finding liars or contradictions
- wherein the at least module is operative to fit, into a fourth fitted label set, the third statistical information using the at least one second learning technique, the at least one second machine learning model, and the at least one second feature set.
[0690]:
In one embodiment, the resulting μ.sub.By (ω) is provided to other modules that take membership function as input (e.g., a fuzzy rule engine) or store in a knowledge data store
[3162]:
the learning process of the learning machine uses a statistical approach to hypothesis parameters by keeping track of those values over time during fitting the data fitting, with more weight given to those parameters occurring more often.
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
[0565]:
The membership function of A, μ.sub.A, may be elicited by asking a succession of questions of the form: To what degree does the number, a, fit your perception of A? Example: To what degree does 50 minutes fit your perception of about 45 minutes? The same applies to B. The fuzzy set, A, may be interpreted as the possibility distribution of X. The concept of a Z-number may be generalized in various ways. In particular, X may be assumed to take values in R.sup.n, in which case A is a Cartesian product of fuzzy numbers.
[1638] :
An approach to approximate or render the overlap (15) between the category sets, such as C.sub.X,n, may use α-cuts to present each crisp α-cuts of predetermined category set as a set of points in (m,σ) space. These sets of points may be modeled efficiently, e.g., based on graphical models, optimized for fast transformation and intersection operations. For example, the models that use peripheral description for the α-cuts allow robust and efficient determination of intersection and avoid the need to transform all the points within the set individually, in order to reduce the computation involved in (13).
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership. As there is overlapping between categories sets (fuzzy sets), it does represent plurality of fitted label sets because an element can have multiple non-zero membership degrees, each indicating the extent to which it fits into a different categories)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
In regard to claim 12: (Original)
McMahan, and Li do not explicitly disclose:
- generating, from a new feature set and the learner unit, a first set of predicted labels.
However, Zadeh discloses:
- generating, from a new feature set and the learner unit, a first set of predicted labels.
[3022]:
detecting or classifying a feature set from said image;
[3023]:
taking an optimization step in training a correlation layer using said feature set and one or more of said invariant or semi-invariant parameters, said variant parameters, and said pose parameters, as input to said correlation layer; and
[3029]:
taking an optimization step in training a correlation layer using said first feature set and said second feature set as input to said correlation layer;
[3031]:
wherein said correlation layer, upon training, outputs a translated feature set, given a third feature set as input to said correlation layer.
(BRI: translated feature set is the new feature)
[3169]:
One embodiment uses predictive feature detection. In one embodiment, the inference module predicts where the features might be based on the initial recognition.
[2782]:
in one embodiment, the indexing uses fuzzy values and intervals. In one embodiment, the search provides one or more potential matching results to a match maker module. In one embodiment, the feature values (fuzzy, crisp, labels) are also provided for coded features (i.e., the features that are descriptive).
[2533]:
In one embodiment, a morph module/application is used to make adjustment to descriptive features/labels after recognizing the model features.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
In regard to claim 13: (Original)
McMahan, and Li do not explicitly disclose:
- querying the at least one module for a second set of predicted labels for the new feature
However, Zadeh discloses:
- querying the at least one module for a second set of predicted labels for the new feature
[2994]:
an input module receiving a first data;
[2983]:
said relevance analysis module generating one or more second relevant items from said one or more first relevant items;
(BRI: a first input module receiving a first item, and a relevance analysis module generating one or more second relevant items from the first relevant items does represent a query in the sense of information retrieval and filtering)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
In regard to claim 14: (Original)
McMahan, and Li do not explicitly disclose:
- repeat the sending and the receiving until and out-sample error no longer decreases
However, Zadeh discloses:
- repeat the sending and the receiving until and out-sample error no longer decreases
[1781]:
In one embodiment, linear models, such as perceptron, linear regression, and/or logistic regression are used. For example, perceptron is used for classification, e.g., in or out, accept or deny, and is trained by perceptron learning algorithm including a pocket version. The linear regression is for example used to determine (continuous valued or real valued) amount, based on squared error function and pseudo-inverse algorithm.
[1781]:
One embodiment uses an adaptive learning rate. In one embodiment, the default learning rate is 0.1. In one embodiment, the number of iterations of epoch is limited to a maximum (early stopping), in order to avoid over fitting the noise/error and deteriorate generalization by increasing the out of sample error.
[0320] :
FIG. 56 shows how to build a fuzzy model, going through iterations, to validate a model, based on some thresholds or conditions.
[1890]:
Of course, even if we get to local minima, rather than absolute minima, for optimization, we still may have a good result for match process, to stop further search and optimization or adjustments, as mentioned above. That can be checked using a relative or absolute value as threshold, or an incremental improvement analysis, to stop beyond a threshold, for the optimization process, as optimization any further would not worth the cost of more computation power spent on such incremental improvements, if any.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
1346]).
In regard to claim 15: (Original)
McMahan does not explicitly disclose:
- the learner unit and the at least one module implement aligned or partially aligned feature datasets.
However, Li discloses:
the learner unit and the at least one module implement aligned or partially aligned feature datasets.
In [Col 23, lines 31-35]:
In accordance with an embodiment and aspect of the invention where gradient data is exchanged, differences in processing rate may cause differences in the amount of data that is evaluated at each device, which may in turn contribute to uncertainty in the gradient calculations.
In [Col 23, lines 41-50]:
In this case, the gradient data may be aggregated with a combining function that weights the different sets of gradient data inversely to a value of the uncertainty metric. In certain cases, the master device may be configured to determine a “samples per second” data processing rate on one or more slave devices. The master device may then dynamically adjust the number of samples per iteration on one or more of the master device and the slave devices to align iteration synchronization with the master device.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, and Li.
McMahan teaches providing user privacy.
Li teaches learning unit.
One of ordinary skill would have motivation to combine McMahan and Li that can improve the versions of neural network model without exchanging any private information (Li [Col 2, lines 54-66])
In regard to claim 16: (Original)
McMahan does not explicitly disclose:
- the processing circuitry is further configured to: limit the at least one module to a particular number based on to at least one of communication bandwidth, cost constraints, or computational overhead.
However, Li discloses:
- the processing circuitry is further configured to: limit the at least one module to a particular number based on to at least one of communication bandwidth, cost constraints, or computational overhead.
in [Col 23, lines 20-30]:
In accordance with an embodiment and aspect of the invention gradient data is exchanged, due to differences between processing speeds on the slave devices and the master device and/or due to differences in the amount of data that each has available for training, there may be large differences in processing time per iteration of configuration data exchange (or per epoch of training). A specific target cycle time may be set (e.g. as a predefined interval) and a back-propagation algorithm on one or more of the master and slave devices may be configured to process enough training data to meet the target cycle time.
In [Col 26, lines 44-49]:
In certain examples, the hyperparameters of the neural network model may be defined according to a predefined standard, and the binary executable may be designed specifically for this standard. This may help to minimize data exchange between the master device and the slave devices, which may speed training iterations.
In [Col 26, lines 39-41]:
Data privacy may also be verified by checking the bandwidth of network traffic between the master device and the slave device.
In [Col 22, lines 63-67], In [Col 23, lines 1-6]:
the master device may be configured to use aggregate data derived from the second configuration data output by the plurality of slave devices to update parameters for the first version of the neural network model. For example, gradient data from the plurality of slave devices may be aggregated by averaging or another suitable statistical function. In one case, the gradient data from different slave devices may be compared to selectively update the parameters for the first version of the neural network model.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, and Li.
McMahan teaches providing user privacy.
Li teaches learning unit.
One of ordinary skill would have motivation to combine McMahan and Li that can improve the versions of neural network model without exchanging any private information (Li [Col 2, lines 54-66])
In regard to claim 17: (Original)
McMahan does not explicitly disclose:
- the learner unit and the at least one module implement centralized feature datasets or decentralized feature datasets.
However, Li discloses:
- the learner unit and the at least one module implement centralized feature datasets or decentralized feature datasets.
In [Col 15, lines 15-23]:
Referring now to FIG. 2, a distributed training system 200 is shown in accordance with an embodiment of the invention. The distributed training system 200 may be seen as a variation of the first distributed training system 100 of FIG. 1 and similar features share similar reference numerals. In accordance with an embodiment and various aspects of the invention, a master device 210 has access to a master data source 264. The master data source 264 may be inaccessible by a slave device 220.
In [Col 15, lines 29-32]:
The master device 210 also has a storage device 260 to store parameter data for an instantiated first version of a neural network model 262. This may be an internal or external storage device.
In [Col 9, lines 43-46]:
each “version” of the neural network model may differ. For example, different versions may include a different number of layers and/or nodes, and/or be configured to process data of differing bit-depths.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, and Li .
McMahan teaches providing user privacy.
Li teaches learning unit.
One of ordinary skill would have motivation to combine McMahan and Li that can improve the versions of neural network model without exchanging any private information (Li [Col 2, lines 54-66])
In regard to claim 19: (Currently Amended)
McMahan discloses:
- A non-transitory, computer-readable medium comprising executable instructions, which when executed by processing circuitry, cause a computing device to perform operations comprising:
In [0005]:
Other example aspects of the present disclosure are directed to systems, apparatus, tangible, non-transitory computer-readable media, user interfaces, memory devices, and electronic devices for solving optimization problems in distributed data environments.
- wherein each of the at least one module executes on at least one remote computing device, and wherein the computing device maintains at least one of algorithm privacy or privacy of the first feature set in sending the first statistical information
In [0004]:
One example aspect of the present disclosure is directed to a computer-implemented method of updating a global model based on unevenly distributed data. The method includes receiving, by one or more computing devices, one or more local updates from a plurality of user devices. Each local update is determined by the respective user device based at least in part on one or more data examples stored on the respective user device. The one or more data examples stored on the plurality of user devices are distributed on an uneven basis, such that no user device includes a representative sample of an overall distribution of data examples. The method further includes aggregating, by the one or more computing devices
in [0054]:
The technology discussed herein makes reference to servers, databases, software applications, and other computer-based systems, as well as actions taken and information sent to and from such systems.
In [0015]:
More particularly, in some embodiments, a set of input-output data can be used to describe a global objective via a loss function. Such functions can be, for instance, a convex or non-convex function, such as a linear regression function, logistic regression function. A local objective (F.sub.k) can also be defined using data stored on a computing device. For instance, the global objective can be defined as:
In [0030]:
FIG. 1 depicts an example system 300 for training one or more global machine learning models 306 using training data 308 stored locally on a plurality of user devices 302. System 300 can further include a server device 304. Server 304 can be configured to access machine learning model 306, and to provide model 306 to a plurality of user devices 302. Model 306 can be, for instance, a linear regression model,
In [0030]:
in some implementations, sever 304 can be configured to communicate with user devices 302 over one or more networks, such as network 240 of FIG. 3.
In [0052]:
The client device 230 can also include a network interface used to communicate with one or more remote computing devices (e.g. server 210) over the network 240.
in [0016]:
According to particular implementations of the present disclosure, the global objective can be solved by aggregating a plurality of local updates provided by a plurality of remote computing devices. Each remote computing device can, for instance, be a user device, such as a laptop computing device, desktop computing device, smartphone, tablet, wearable computing device, or other computing device. The local updates can be determined based at least in part on the respective local objectives.
in [0054]:
Databases and applications may be implemented on a single system or distributed across multiple systems.
in [0014]:
the local update may be determined using one or more gradient descent techniques
in [0014]:
the local update does not include the training data used to determine the local update. In this manner, the size of the local update can be independent of the training data used to determine the local update, thereby reducing bandwidth requirements and maintaining user privacy. In particular, a global model can be updated based at least in part on the received local updates. By only providing the local update (and not the training data) to the server, the global model update can be determined using reduced bandwidth requirements, and without compromising the security of potentially privacy sensitive data stored on the user devices
McMahan does not explicitly disclose:
- creating a learner unit by fitting, into a first fitted label set, an initial label set using at least one first learning technique, a first machine learning model, and a first feature set;
- sending, to at least one module in a machine learning architecture, first statistical information defined by at least one first residual from fitting the first fitted label set and excluding the first feature set and excluding the first feature set, wherein the at least one module is operative to fit, into a second fitted label set,
- the first statistical information using at least one second learning technique, at least one second machine learning model, and at least one second feature set,
- receiving, from the at least one module, second statistical information that is defined by at least one second residual from fitting the second fitted label set and excluding the second feature set
- and updating the learner unit by fitting, into a third fitted label set, the second statistical information using the at least one first learning technique and the first machine learning model.
However, Li discloses:
- a learner unit by fitting, into a first fitted label set
In [Col 4, lines 41-56]:
In accordance with various aspects of the present invention, the first configuration data includes hyperparameters for the neural network model and parameters for the first version of the neural network model. The hyperparameters may include one or more of: an architecture definition for the neural network model; a number of nodes for one or more layers in the neural network model; a set of node definitions indicating one or more of a node type and a node connectivity; a set of activation function definitions; and one or more cost function definitions. The parameters may include one or more of: weight values for one or more connections between nodes of the neural network model; weight values for one or more inputs to the neural network model; weight values for one or more recurrent paths in the neural network model; and bias values for one or more nodes of the neural network model.
(BRI: A learner unit is basic computational element that performs a calculation on its inputs to produce an output and passing the result through an activation function, enabling the network to learn complex patterns in data during the training process)
In [ Col 9, lines 17-23]:
“neural network model” is used to refer to an artificial neural network that is configured to perform a particular data processing task. For example, in the case that a neural network model includes an acoustic model, the task may be to output phoneme or grapheme data (e.g. predictions of phonemes or graphemes) based on input audio data.
In [ Col 9, lines 26-32]:
In certain cases, a neural network model may be a model that is configured to provide a particular mapping between defined input data and defined output data. The input data may represent one modality and output data may represent another modality. The neural network model may be considered a function approximator that is trained on a set of data.
(BRI: A "fitted label set" is a concept most relevant to machine learning, where it refers to the mapping of numerical labels to categorical data after a model has been "fitted" or trained. A modality can act as a type of "label" or category to describe a type of data)
In [Col 19, lines 58-62]:
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570.
- an initial label set using at least one first learning technique, a first machine learning model, and a first feature set;
In [Col 12, lines 2-6]:
a version of the neural network model may be instantiated by implementing a class or class-like definition of the neural network model using initialization data. In this case, the initialization data includes the first configuration data 180.
- sending, by the processing circuitry of the computing device, to at least one module in a machine learning architecture,
In [Col 22, lines 49-53]:
a loss function may be a function of log it values across the ensemble of instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.)
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating the parameter data for the first version of the neural network model based on the second configuration data.
- receiving, by the processing circuitry of the computing device, and from the at least one module,
In [Col 22, lines 49-53]:
a loss function may be a function of log it values across the ensemble of instantiations of the second version of the neural network model. A loss function may aggregate the log it values from the ensemble, e.g. via averaging or another statistical computation (median, mode etc.)
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating the parameter data for the first version of the neural network model based on the second configuration data.
- and updating the learner unit by fitting, into a third fitted label set, the second statistical information using the at least one first learning technique and the first machine learning model.
In [Col 19, lines 58-67], In [Col 20, lines 1-3] :
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570. As per FIG. 4A or 4B, each of the versions output log its that are compared in a loss function to steer the gradient descent update of the parameters of the first version of the neural network model 562. Following training in this manner, the parameter data in the data storage device 560 is updated, allowing a revised or updated set of first configuration data 580 to be generated. This can then be communicated to a set of slave devices to repeat the process.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, and Li.
McMahan teaches providing user privacy.
Li teaches data set of fit, label and plurality of models and a learning unit.
One of ordinary skill would have motivation to combine McMahan and Li that can improve the versions of neural network model without exchanging any private information (Li [Col 2, lines 54-66])
However, Zadeh discloses:
- first statistical information that includes at least one first residual from fitting the first fitted label set that excludes both the first feature set, and first model information of the first machine model, from disclosure to the at least one module, wherein module is operative to fit, into at least one second fitted label set
[3162]:
the learning process of the learning machine uses a statistical approach to hypothesis parameters by keeping track of those values over time during fitting the data fitting, with more weight given to those parameters occurring more often.
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
[0565]:
The membership function of A, μ.sub.A, may be elicited by asking a succession of questions of the form: To what degree does the number, a, fit your perception of A? Example: To what degree does 50 minutes fit your perception of about 45 minutes? The same applies to B. The fuzzy set, A, may be interpreted as the possibility distribution of X. The concept of a Z-number may be generalized in various ways. In particular, X may be assumed to take values in R.sup.n, in which case A is a Cartesian product of fuzzy numbers.
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership, the set is the fitted labeled set)
[2129]:
For the aggregation method (also called ensemble learning, or boosting, or mixture of experts), we have a learning which tries to replicate the function independently (not jointly), and then combine and put them together later,
[2129]:
For the aggregation method, for regression or real number cases, we take an average or weighted average,
[2129]:
For the aggregation method, we have 2 types: (a) After-the-fact situation (where we already have the solutions, and then we combine them), and (b) Before-the-fact situation (where we get solutions, with the view or intention or assumption to blend or combine them together later). For the aggregation method, as one example, we have the Boosting method, where we enforce the decorrelation (not by chance), e.g. by building one hypothesis at a time, for a good mixture, sequentially.
[2092] :
In one embodiment, for aggregating the correlated fuzzy sets, for the Min-Max aggregation method, we can get the final membership value, μ.sub.final, based on the individual membership values, μ.sub.i and μ.sub.2, as (where index i runs from 0 to n):
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
[1638]:
An approach to approximate or render the overlap (15) between the category sets, such as C.sub.X,n, may use α-cuts to present each crisp α-cuts of predetermined category set as a set of points in (m,σ) space. These sets of points may be modeled efficiently, e.g., based on graphical models, optimized for fast transformation and intersection operations. For example, the models that use peripheral description for the α-cuts allow robust and efficient determination of intersection and avoid the need to transform all the points within the set individually, in order to reduce the computation involved in (13).
(BRI: Elements and their degree of membership and within the context of A and B as membership. As there is overlapping between categories sets (fuzzy sets), it does represent plurality of fitted label sets because an element can have multiple non-zero membership degrees, each indicating the extent to which it fits into a different categories)
[1723]:
In one embodiment, the added units and the previous units are used to make association and/or correlation with labeled samples, e.g., during the supervised training.
[1723]:
In one embodiment, the labels are continuous valued (or multi-valued), e.g., having values in range [0, 1], to indicate the degree in which the document is classified by a label (or the membership function of the document in the label's class).
[0113]:
For one embodiment: Decisions are based on information. To be useful, information must be reliable. Basically, the concept of a Z-number relates to the issue of reliability of information. A Z-number, Z, has two components, Z=(A,B). The first component, A, is a restriction (constraint) on the values which a real-valued uncertain variable, X, is allowed to take. The second component, B, is a measure of reliability (certainty) of the first component.
[0564]:
For purposes of computation, when A and B are described in a natural language, the meaning of A and B is precisiated (graduated) through association with membership functions, μ.sub.A and μ.sub.B, respectively, FIG. 1.
[0593]:
A basic fuzzy if-then rule may be expressed as: if X is A then Y is B, where A and B are fuzzy numbers. The meaning of such a rule is defined as:
PNG
media_image2.png
17
202
media_image2.png
Greyscale
[1678] :
In one embodiment, the learning is achieved through simulation using a data (and label) sample generation based on one or more models. In one embodiment, a network trained based on model(s) is used to recognize and classify actual data which may not have been seen before. In one embodiment, the system is trained to infer the potential model(s) itself by recognizing the (e.g., observed) data conforming to a particular model and its associated labels/parameters.
[1757]:
The features of both images (and in particular the differences between their features) are correlated using a correlator/analyzer module (e.g., having unit/neurons) with the label/feature difference identified in the label layer (e.g., L.sub.1). In one embodiment, the L layer represents the labels indicated the feature differences between the images (denoted by ΔL). In one embodiment, more than one label is selected by the controller (indicating the differences between the features of the images selected from the database for training the correlation).
[2431]:
In one embodiment, the system uses statistics and patterns extracted from images
[2788]:
a script is executed that iterate through the document object model or the window object to identify elements associated with images
[1723]:
In one embodiment, feature detection system is used to train document classification based on learned (e.g., unsupervised) features corresponding to documents based on terms contained in the document (such as statistics of several hundred or several thousand common words)
[2521]:
In one embodiment, as for example depicted in FIG. 228, based on an identifier or a URL, a movie and its associated data (e.g., metadata, title, description, owner/uploader, channel, comments, likes, and statistics) are extracted from a repository. In one embodiment, the movie and its associated data are analyzed, e.g., by a video/audio analyzer and keyword/concept extraction/analyzer, to provide/generate features/annotation and metadata
[1757]:
In one embodiment, the weights related to the units in the correlator/analyzer are trained to detect the feature differences by a stochastic or batch learning algorithm.
[1734]:
features of an object (e.g., pose including rotation) is determined, and based on such features, features of sub-objects of other objects depicted in an image are extracted by preprocessing (e.g., mapping) a portion of an image into a segmented layout with variable resolution. Then, the mapped image (or portion thereof) is provided to a classifier or feature recognition system to determine the features from the mapped image
[1351]:
In one embodiment, an optimum statistical classifier is used. In one embodiment, a Bayes classifier is used
[1351]:
one embodiment, a perceptron for 2-pattern classes is used. In one embodiment, the least mean square (LMS) delta rule for training perceptrons is used, to minimize the error between the actual response and the desired response (for the training purposes). FIG. 115 is an example of a system described above.
[1352] :
In one embodiment, a multi-layer feed-forward neural network is used. In one embodiment, the training is done by back propagation, using the total squared error between the actual responses and desired responses for the nodes in the output layer.
[2301]:
In one embodiment, we segment the data of any type, including video, sound, and multimedia, based on sudden change in the sequence (or big delta or difference),
[2300]:
In one embodiment, we use Bayesian model, for both sides of the potential boundary between segments, with 2 different model parameters, to fit the 2 sides better, to examine the potential boundary for segmentation, e.g. for speech.
(BRI: in the context of a neural network or any predictive model, the difference between the actual output (response) and the desired output (target) for the nodes in the output layer is indeed the residual. From a statistical information perspective, these residuals are not just raw differences and they are statistical information in the sense that they quantify the discrepancy between the model’s predictions and the true data. The plurality of residuals are within the context of residual associated to each node and the statistical information associated to each residual)
[1765]:
to detect the object/class of object based on a thumbnail (e.g., via preprocessing) of the data/image. In one embodiment, the training for a high level feature detection focuses on the structure of the neurons or units used in a classifier/feature detector. In one embodiment, the resulting feature units at top layer are limited to few features, while the training is used with data/images that may include thumbnail and high resolution data/images, including those with and without the targeted features.
[1768] :
In one embodiment, meta data such as the GPS data (or for example other accompanying metadata captured with images taken from mobile devices such as smart phones) are used as labels (e.g., continuous valued).
[1586]:
The communication between different units, devices, or modules are done by wire, cable, fiber optics, wirelessly, WiFi, Bluetooth, through network, Internet, copper interconnect, antenna, satellite dish, or the like.
1759]:
In one embodiment, one or more pose detection modules (e.g., based on edge detection or color region/shape) are used to determine the pose of a face within an image/data
[1810]:
FIG. 127 shows a system for context determination, with language input device, which feeds dissecting and parsing modules to get the components or parts of the sentence, which feeds the analyzing module (which e.g. may include memory units and processor units or CPU or computing module), which is connected to the context determination module, which is connected to the default analyzer module and multiple other context analyzer module
PNG
media_image3.png
537
513
media_image3.png
Greyscale
[2422]:
In one embodiment, we have private or public or semi-public or semi-private (or the like) settings for our sharing or displaying or reviewing or tagging or annotating or accessing or searching or browsing of images or objects or videos, for user, friends, family, co-worker, boss, employee, contractor, senior management, public, social network, college, school, classmate, roommate, household, shared device, shared account, friend-of-friend, friend-of-friend-of-friend, and so on, or the like. In one embodiment, we have government excluding list database, for specified individuals, to exclude for the rules, for the above functions, for privacy settings. In one embodiment, we have the intersection of the privacy settings of multiple users or contributors
[2789]:
In one embodiment, the script determines whether to identify the images based on the domain name (e.g., as positive filter to include or negative filter to exclude).
[2830]:
publisher's restriction/filters to exclude certain or certain types of merchants and merchant's restriction/filters to exclude certain or certain types of publishers.
[1917]:
In one embodiment, for conditional relationships, or multiple choices, we can continue, until we get to a dead end or conflict, and then, backtrack to eliminate or adjust one or more choices, on the chain going backward, to correct or adjust some assumptions, choices, or conditions, on the way.
[2355]:
we can input these conditions or rules into our rule engine, or use it for prediction, control system, forecasting (economy, elections, and other events), social behavioral analysis, consumer behavioral analysis, predicting revolutions or unrest, detecting frauds, detecting unusual behaviors, detecting unusual patterns, finding liars or contradictions
- second statistical information first statistical information that includes at least one second residual from fitting the second fitted label set that excludes both the second feature set, and second model information of the second machine model, and that is based on the first statistical information, from disclosure to the at least one module, and that is based on the first statistical information;
[3162]:
the learning process of the learning machine uses a statistical approach to hypothesis parameters by keeping track of those values over time during fitting the data fitting, with more weight given to those parameters occurring more often.
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
[0565]:
The membership function of A, μ.sub.A, may be elicited by asking a succession of questions of the form: To what degree does the number, a, fit your perception of A? Example: To what degree does 50 minutes fit your perception of about 45 minutes? The same applies to B. The fuzzy set, A, may be interpreted as the possibility distribution of X. The concept of a Z-number may be generalized in various ways. In particular, X may be assumed to take values in R.sup.n, in which case A is a Cartesian product of fuzzy numbers.
[1638]:
An approach to approximate or render the overlap (15) between the category sets, such as C.sub.X,n, may use α-cuts to present each crisp α-cuts of predetermined category set as a set of points in (m,σ) space. These sets of points may be modeled efficiently, e.g., based on graphical models, optimized for fast transformation and intersection operations. For example, the models that use peripheral description for the α-cuts allow robust and efficient determination of intersection and avoid the need to transform all the points within the set individually, in order to reduce the computation involved in (13).
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership, the set is the fitted labeled set, and the elements and their degree of membership and within the context of A and B as membership. As there is overlapping between categories sets (fuzzy sets), it does represent plurality of fitted label sets because an element can have multiple non-zero membership degrees, each indicating the extent to which it fits into a different categories)
[2129]:
For the aggregation method (also called ensemble learning, or boosting, or mixture of experts), we have a learning which tries to replicate the function independently (not jointly), and then combine and put them together later,
[2129]:
For the aggregation method, for regression or real number cases, we take an average or weighted average,
[2129]:
For the aggregation method, we have 2 types: (a) After-the-fact situation (where we already have the solutions, and then we combine them), and (b) Before-the-fact situation (where we get solutions, with the view or intention or assumption to blend or combine them together later). For the aggregation method, as one example, we have the Boosting method, where we enforce the decorrelation (not by chance), e.g. by building one hypothesis at a time, for a good mixture, sequentially.
[2092] :
In one embodiment, for aggregating the correlated fuzzy sets, for the Min-Max aggregation method, we can get the final membership value, μ.sub.final, based on the individual membership values, μ.sub.i and μ.sub.2, as (where index i runs from 0 to n):
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership, the set is the fitted labeled set)
[1723]:
In one embodiment, the added units and the previous units are used to make association and/or correlation with labeled samples, e.g., during the supervised training.
[1723]:
In one embodiment, the labels are continuous valued (or multi-valued), e.g., having values in range [0, 1], to indicate the degree in which the document is classified by a label (or the membership function of the document in the label's class).
[0113]:
For one embodiment: Decisions are based on information. To be useful, information must be reliable. Basically, the concept of a Z-number relates to the issue of reliability of information. A Z-number, Z, has two components, Z=(A,B). The first component, A, is a restriction (constraint) on the values which a real-valued uncertain variable, X, is allowed to take. The second component, B, is a measure of reliability (certainty) of the first component.
[0564]:
For purposes of computation, when A and B are described in a natural language, the meaning of A and B is precisiated (graduated) through association with membership functions, μ.sub.A and μ.sub.B, respectively, FIG. 1.
[0593]:
A basic fuzzy if-then rule may be expressed as: if X is A then Y is B, where A and B are fuzzy numbers. The meaning of such a rule is defined as:
PNG
media_image2.png
17
202
media_image2.png
Greyscale
[1678] :
In one embodiment, the learning is achieved through simulation using a data (and label) sample generation based on one or more models. In one embodiment, a network trained based on model(s) is used to recognize and classify actual data which may not have been seen before. In one embodiment, the system is trained to infer the potential model(s) itself by recognizing the (e.g., observed) data conforming to a particular model and its associated labels/parameters.
[1757]:
The features of both images (and in particular the differences between their features) are correlated using a correlator/analyzer module (e.g., having unit/neurons) with the label/feature difference identified in the label layer (e.g., L.sub.1). In one embodiment, the L layer represents the labels indicated the feature differences between the images (denoted by ΔL). In one embodiment, more than one label is selected by the controller (indicating the differences between the features of the images selected from the database for training the correlation).
[2431]:
In one embodiment, the system uses statistics and patterns extracted from images
[2788]:
a script is executed that iterate through the document object model or the window object to identify elements associated with images
[1723]:
In one embodiment, feature detection system is used to train document classification based on learned (e.g., unsupervised) features corresponding to documents based on terms contained in the document (such as statistics of several hundred or several thousand common words)
[2521]:
In one embodiment, as for example depicted in FIG. 228, based on an identifier or a URL, a movie and its associated data (e.g., metadata, title, description, owner/uploader, channel, comments, likes, and statistics) are extracted from a repository. In one embodiment, the movie and its associated data are analyzed, e.g., by a video/audio analyzer and keyword/concept extraction/analyzer, to provide/generate features/annotation and metadata
[1757]:
In one embodiment, the weights related to the units in the correlator/analyzer are trained to detect the feature differences by a stochastic or batch learning algorithm.
[1734]:
features of an object (e.g., pose including rotation) is determined, and based on such features, features of sub-objects of other objects depicted in an image are extracted by preprocessing (e.g., mapping) a portion of an image into a segmented layout with variable resolution. Then, the mapped image (or portion thereof) is provided to a classifier or feature recognition system to determine the features from the mapped image
[1351]:
In one embodiment, an optimum statistical classifier is used. In one embodiment, a Bayes classifier is used
[1351]:
one embodiment, a perceptron for 2-pattern classes is used. In one embodiment, the least mean square (LMS) delta rule for training perceptrons is used, to minimize the error between the actual response and the desired response (for the training purposes). FIG. 115 is an example of a system described above.
[1352] :
In one embodiment, a multi-layer feed-forward neural network is used. In one embodiment, the training is done by back propagation, using the total squared error between the actual responses and desired responses for the nodes in the output layer.
[2301]:
In one embodiment, we segment the data of any type, including video, sound, and multimedia, based on sudden change in the sequence (or big delta or difference),
[2300]:
In one embodiment, we use Bayesian model, for both sides of the potential boundary between segments, with 2 different model parameters, to fit the 2 sides better, to examine the potential boundary for segmentation, e.g. for speech.
(BRI: in the context of a neural network or any predictive model, the difference between the actual output (response) and the desired output (target) for the nodes in the output layer is indeed the residual. From a statistical information perspective, these residuals are not just raw differences and they are statistical information in the sense that they quantify the discrepancy between the model’s predictions and the true data. The plurality of residuals are within the context of residual associated to each node and the statistical information associated to each residual)
[1765]:
to detect the object/class of object based on a thumbnail (e.g., via preprocessing) of the data/image. In one embodiment, the training for a high level feature detection focuses on the structure of the neurons or units used in a classifier/feature detector. In one embodiment, the resulting feature units at top layer are limited to few features, while the training is used with data/images that may include thumbnail and high resolution data/images, including those with and without the targeted features.
[1768] :
In one embodiment, meta data such as the GPS data (or for example other accompanying metadata captured with images taken from mobile devices such as smart phones) are used as labels (e.g., continuous valued).
[1586]:
The communication between different units, devices, or modules are done by wire, cable, fiber optics, wirelessly, WiFi, Bluetooth, through network, Internet, copper interconnect, antenna, satellite dish, or the like.
1759]:
In one embodiment, one or more pose detection modules (e.g., based on edge detection or color region/shape) are used to determine the pose of a face within an image/data
[1810]:
FIG. 127 shows a system for context determination, with language input device, which feeds dissecting and parsing modules to get the components or parts of the sentence, which feeds the analyzing module (which e.g. may include memory units and processor units or CPU or computing module), which is connected to the context determination module, which is connected to the default analyzer module and multiple other context analyzer module
PNG
media_image3.png
537
513
media_image3.png
Greyscale
[2422]:
In one embodiment, we have private or public or semi-public or semi-private (or the like) settings for our sharing or displaying or reviewing or tagging or annotating or accessing or searching or browsing of images or objects or videos, for user, friends, family, co-worker, boss, employee, contractor, senior management, public, social network, college, school, classmate, roommate, household, shared device, shared account, friend-of-friend, friend-of-friend-of-friend, and so on, or the like. In one embodiment, we have government excluding list database, for specified individuals, to exclude for the rules, for the above functions, for privacy settings. In one embodiment, we have the intersection of the privacy settings of multiple users or contributors
[2789]:
In one embodiment, the script determines whether to identify the images based on the domain name (e.g., as positive filter to include or negative filter to exclude).
[2830]:
publisher's restriction/filters to exclude certain or certain types of merchants and merchant's restriction/filters to exclude certain or certain types of publishers.
[1917]:
In one embodiment, for conditional relationships, or multiple choices, we can continue, until we get to a dead end or conflict, and then, backtrack to eliminate or adjust one or more choices, on the chain going backward, to correct or adjust some assumptions, choices, or conditions, on the way.
[2355]:
we can input these conditions or rules into our rule engine, or use it for prediction, control system, forecasting (economy, elections, and other events), social behavioral analysis, consumer behavioral analysis, predicting revolutions or unrest, detecting frauds, detecting unusual behaviors, detecting unusual patterns, finding liars or contradictions
- updating the learner unit by fitting, into a third fitted label set, the second statistical information using the at least one first learning technique and the first machine learning model.
[1340]:
the fuzzy classifier module or device classifies or separates different pictures into clusters or groups in N-dimensional feature space.
[1531]:
a method for fuzzy logic control, in which an input module receives a precisiated proposition associated with a protoform. A fuzzy logic inference engine evaluates a first fuzzy logic rule from the fuzzy logic rule repository. The fuzzy logic inference engine is in or loaded on or executed on or implemented in a computing device, which comprises one or more of following: computer, processor device
[0565]:
The membership function of A, μ.sub.A, may be elicited by asking a succession of questions of the form: To what degree does the number, a, fit your perception of A? Example: To what degree does 50 minutes fit your perception of about 45 minutes? The same applies to B. The fuzzy set, A, may be interpreted as the possibility distribution of X. The concept of a Z-number may be generalized in various ways. In particular, X may be assumed to take values in R.sup.n, in which case A is a Cartesian product of fuzzy numbers. Simple examples of Z-valuations are:
(BRI: When fitting a model to data, the goal is to find parameters that best match observed values. With fuzzy data, the “fit” is adjusted to account for uncertainty)
[0907]:
the fuzzy set labeled small is the possibility distribution of X. If μ. sub. small is the membership function of small, then the semantics of “X is small” is defined by
PNG
media_image1.png
42
431
media_image1.png
Greyscale
[0908] where u is a generic value of X.
[0909] (b) Probabilistic (r=p)
X isp R,
(BRI: the membership function itself is a representation of the label set that encodes both the set of elements and their degree of membership and within the context of A and B as membership, the set is the fitted labeled set)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
In regard to claim 20: (Original)
- generating, from a new feature set and the learner unit, a first set of predicted labels;
[3022]:
detecting or classifying a feature set from said image;
[3023]:
taking an optimization step in training a correlation layer using said feature set and one or more of said invariant or semi-invariant parameters, said variant parameters, and said pose parameters, as input to said correlation layer; and
[3029]:
taking an optimization step in training a correlation layer using said first feature set and said second feature set as input to said correlation layer;
[3031]:
wherein said correlation layer, upon training, outputs a translated feature set, given a third feature set as input to said correlation layer.
(BRI: translated feature set is the new feature)
3169]:
One embodiment uses predictive feature detection. In one embodiment, the inference module predicts where the features might be based on the initial recognition.
[2782]:
in one embodiment, the indexing uses fuzzy values and intervals. In one embodiment, the search provides one or more potential matching results to a match maker module. In one embodiment, the feature values (fuzzy, crisp, labels) are also provided for coded features (i.e., the features that are descriptive).
[2533]:
In one embodiment, a morph module/application is used to make adjustment to descriptive features/labels after recognizing the model features.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
- querying the at least one module for a second set of predicted labels for the new feature set;
[2994]:
an input module receiving a first data;
[2983]:
said relevance analysis module generating one or more second relevant items from said one or more first relevant items;
(BRI: a first input module receiving a first item, and a relevance analysis module generating one or more second relevant items from the first relevant items does represent a query in the sense of information retrieval and filtering)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
- combining the first set of predicted labels and the second set of predicted labels.
[2770]:
The features representing the object/person/face are combined or enhanced based on the reliability of the features from different sources.
[1739]:
In one embodiment, the locations of interest (e.g., the location of faces within an image) is determined by a scanning the image through a variable size window over an image at different location on the image, searching for example for particular features or signatures (e.g., head or face).
[1750]:
In one embodiment, classifiers are trained to detect high level signatures/features of various objects/concepts, e.g., by training the classifiers with (labeled) training data sets,
[1739]:
such image and the associated information are used to train a feature detector/classifier to learn or predict the focuses of interest, by correlating/associating the image features with the locations of interest. In one embodiment, the image and various positions of interest are iteratively inputted to the system during training. The stochastic nature of the correlation layer, stochastically reconstruct parameters associated with the location of interest as output, e.g., using an RBM.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li and Zadeh.
McMahan teaches providing user privacy.
Li teaches learner unit.
Zadeh teaches residual and prediction and exclusion of feature sets.
One of ordinary skill would have motivation to combine McMahan, Li and Zadeh can provide optimum performance for example for autonomous driving cars (Zadeh[1558])
Claim 18 is rejected under 35 U.S.C. 103 as being unpatentable over
Hugh McMahan (hereinafter McMahan) US 2017/0109322 A1,
in view of Zili Li et.al. (hereinafter Li) US 11551083 B2,
further in view of Eugene Tuv et.al. (hereinafter Tuv) Feature Selection with Ensembles, Artificial Variables, and Redundancy Elimination, Journal of Machine Learning Research 10 (2009) 1341-1366.
In regard to claim 18: (Original)
McMahan does not explicitly disclose:
- train the first machine learning model using the at least one first learning technique with the initial label set and the first feature set, wherein the trained machine learning model is configured to generate the first fitted label set for the first feature set;
- wherein to send, to the at least one module in the machine learning architecture,
- determine a first particular residual of the at least one first residual based on a first fitted label of the first fitted label set and an observed data set in the first feature set, wherein the at least one module trains the least one second machine learning model using the at least one second learning technique with the at least one second feature set, wherein a second particular residual of the least one second residual is determined from a second fitted label of the second fitted label set and the observed data set;
- further train the trained machine learning model with the at least one second residual and the first feature set.
However, Li discloses:
- train the first machine learning model using the at least one first learning technique with the initial label set and the first feature set, wherein the trained machine learning model is configured to generate the first fitted label set for the first feature set;
In [Col 19, lines 58-62]:
During training, which again may be on one or more of unlabeled data and labelled
data, the parameters of the first version of the neural network model 562 are updated based on processing by the second version of the neural network model 570.
(BRI: fitting" a model to data is a key part of the training process in machine learning, where the model learns to make accurate predictions by adjusting its parameters based on the provided data).
- wherein to send, to the at least one module in the machine learning architecture,
In [Col 27, lines 15-28]:
The master device may use the network interface to communicate with one or more slave devices. The at least one processor of the master device may be configured to execute computer program code stored in memory to perform one or more operations. These operations may include: generating first configuration data for the neural network model based on the first version of the neural network model; sending, via the network interface, the first configuration data to the slave device 920; receiving, via the network interface, second configuration data 990 for the neural network model from the slave device; and updating the parameter data for the first version of the neural network model based on the second configuration data.
McMahan and Li do not explicitly disclose:
- determine a first particular residual of the at least one first residual based on a first fitted label of the first fitted label set and an observed data set in the first feature set, wherein the at least one module trains the least one second machine learning model using the at least one second learning technique with the at least one second feature set, wherein a second particular residual of the least one second residual is determined from a second fitted label of the second fitted label set and the observed data set;
- further train the trained machine learning model with the at least one second residual and the first feature set.
However, Tuv discloses:
- determine a first particular residual of the at least one first residual based on a first fitted label of the first fitted label set and an observed data set in the first feature set, wherein the at least one module trains the least one second machine learning model using the at least one second learning technique with the at least one second feature set, wherein a second particular residual of the least one second residual is determined from a second fitted label of the second fitted label set and the observed data set;
In [4.1, 1351]:
The Gram-Schmidt procedure first selects the variable with highest correlation with the target. To remove the information from this variable the remaining predictors and the target are orthogonalized with respect to the selected variable. This provides residuals from the fit of the target to the first selected variable
(BRI: "orthogonalized with a variable" in a statistical context adjust a variable to make it uncorrelated with one or more other variables, effectively removing the shared information or overlap between them. The first selected variable is the first feature set).
In [3, Page 1346]:
Supervised ensemble methods construct a set of simple models, called base learners, and use their weighted outcome (or vote) to predict new data. That is, ensemble methods combine outputs from multiple base learners to form a committee with improved performance.
In[ 4.2, Page 1354]:
The learner g(.,.) in the ACE algorithms is an ensemble of trees. Any classifier/regressor function can be used, from which the variable importance from all variable interactions can be derived.
In[ 4.2, Page 1353]:
Our ACE algorithm works more like FCBS as it uses only one feature as an approximate MB for each feature (as does the MBBE algorithm with K = 1). Furthermore, it filters features by relevance before computing redundancy between the features, and reports a final minimum feature subset.
In [4.1 Algorithm Detail, Page 1349]:
1. Identify Important Variables: Artificially generated noise variables are used to determine a threshold to test for statistically significant variable importance scores. The test is used to remove irrelevant variables. Details are presented in the displayed algorithms and further described as follows. In each replicate r, r = 1,2,...,R artificial variables are constructed as follows. For every real variable Xj j = 1,2,...,M a corresponding artificial variable Zj is generated from a random permutation. Then in each replicate a small RF is trained and variable importance scores are computed for real and artificial variables.
In [4.1 Algorithm Detail, Page 1350]:
A test that results in statistical significance identifies an important variable.
In [4.1 Algorithm Detail, Page 1350]:
4. Generate Residuals for Incremental Adjustment: An iteration is used to enhance the ability of the algorithm to detect variables that are important,
[4.1 Algorithm Detail, Page 1350]:
Given a current subset of important variables, only this subset is used to predict the target. Residuals are calculated and form a new target. For a numerical target the residuals are simply the actual minus the predicted values.
The algorithms for numerical (regression) and categorical (classification) targets are presented as algorithms 1 and 2. A separate algorithm 3 describes the variable masking calculations.
[4.1 Algorithm Detail, Page 1351]:
PNG
media_image4.png
685
1050
media_image4.png
Greyscale
[4.1 Algorithm Detail, Page 1351]:
PNG
media_image5.png
492
592
media_image5.png
Greyscale
In [4.2, Page 1352]:
FCBS first sorts features by correlation with the response using a symmetric
feature most correlated to the response is selected. 2. All features that have correlation with the selected feature higher than it’s correlation with response are considered redundant and removed. The feature is added to the minimal subset (and this is an approximate heuristic for Markov blanket filtering). 3. Return to 1
In [4.2, Page 1354]:
(BRI: within the context of using notations for algorithms and algorithm 2, the plurality of statistical information is used (from residual) for label fitting) and the exclusion as provided based on the comparing to a significance is provided)
- further train the trained machine learning model with the at least one second residual and the first feature set.
In [2.2, Page 1344]:
linear regression with backward feature elimination are heavily dependent on the model (linear or SVM), that can fail to fit the data well.
In [2.2, Page 1344]:
Data sets with tens of thousands of features or samples become very time consuming and impractical to handle. For example, SVM-RFE involves retraining the SVM after features with smallest relevance are removed
In [3, Page 1346]:
At every step of ensemble construction the boosting scheme adds a new base learner that is forced (by iteratively reweighting the training data) to concentrate on the training observations that are misclassified by the previous sequence
4.1 Algorithm Detail, Page 1351]:
PNG
media_image5.png
492
592
media_image5.png
Greyscale
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine McMahan, Li , Zadeh nd Tuv.
McMahan teaches providing user privacy.
Li teaches train the ML model.
Tuv teaches a particular residual and further training.
One of ordinary skill would have motivation to combine McMahan, Li , Zadeh and Tuv that can improve the performance using multiple learners (Tuv [3, Page 1346]).
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the
examiner should be directed to TIRUMALE KRISHNASWAMY RAMESH whose telephone number is (571)272-4605. The examiner can normally be reached by phone.
Examiner interviews are available via telephone, in-person, and video conferencing
using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at
http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Li B Zhen can be reached on phone (571-272-3768). The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be
obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit:
https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for
information about filing in DOCX format.
For additional questions, contact the Electronic
Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO
Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/TIRUMALE K RAMESH/Examiner, Art Unit 2121
/Li B. Zhen/Supervisory Patent Examiner, Art Unit 2121
Prosecution Insights