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 .
Examiner notes the entry of the following papers:
Amended claims filed 11/7/2025.
Applicant arguments/remarks mad in amendment filed 11/7/2025.
Claims 1-2, 8-9, and 15-16 are amended. Claims 19-20 are cancelled. Claims 21-22 are new. Claims 1-18, and 21-22 are presented for examination.
Response to Arguments
Applicant presents arguments. Each is addressed.
Applicant argues “Amended claim 1 reflects ‘an improvement’ to a problem in computer technology such as ‘models are often limited due to their complex parameterization relying on discrete time steps’. (Specification, [0017] The claimed invention in amended claim 1 ‘improve the efficiency of time-series forecasting….” (Remarks, page 9, paragraph 3, line 1). However, 2106.05(a) recites "It is important to note, the judicial exception alone cannot provide the improvement. The improvement can be provided by one or more additional elements." Applicant has not provided objective evidence of a technical improvement but is instead relying entirely on the judicial exception to provide a technical improvement. Therefore, the rejection is maintained.
Applicant argues “Specifically, the Judge also cited to McRo and reasoned the claim at issue "recites the use of specific mathematical operations," and "the specification indicates leads to faster training and improved operation of the neural network," which is "adequate to show that the claims recite a particular method that reflects an improvement to existing technology." (Decision on Appeal, pp. 6-7). Therefore, under similar rationale, the amended independent claims, as discussed above, also recites the use of alleged mathematical operations such as "training the neural network by updating the first parameters of the final layer based on a first training objective" and "training the neural network by updating the second parameters based on a second training objective".” (Remarks, page 10, paragraph 2, line 7.) However, the claimed invention does not claim to show faster training. Nor is any objective evidence provided to support that claim. Therefore, the analogy fails and the rejection is maintained.
Applicant argues “For example, Applicant respectfully points the Examiner to the latest decision in Appeal 2024-000567, in which Director John A Squires decided a method claim of "training a machine learning model" is eligible. (Decision on Appeal 2024-000567, App. 16/319,040)…. Applicant respectfully submits the pending claims, which reflect an improvement to time-series forecasting neural network models/systems as described in the specification, are thus eligible subject matter.” (Remarks, page 10, paragraph 3, line 1). However, as mentioned above, 2106.05(a) recites "It is important to note, the judicial exception alone cannot provide the improvement. The improvement can be provided by one or more additional elements." Applicant has not provided objective evidence of a technical improvement but is instead relying entirely on the judicial exception to provide a technical improvement. Therefore, the rejection is maintained.
Applicant remarks “Claims 2, 9, and 16 have been amended to add proper punctuation to end the claim. Applicant respectfully requests the reconsideration and withdrawal of the objection.” (Remarks, page 11, paragraph 4, line 1) Examiner notes the amendment. The objections to claims 2, 9, and 16 have been withdrawn.
Applicant argues that “For example, the as-filed Specification describes that "the training of the ridge regressor may be considered an inner-loop optimization to minimize a first training objective while updating the parameters of the ridge regressor, which is done between outer-loop optimizations of gΘ, which minimizes a second training objective while updating the parameters of gΘ with the parameters of the ridge regressor temporarily frozen" (Remarks, page 11, paragraph 7, line 1.) However, the claims do not recite that. There is no mention of “ridge regressor”, “inner-loop”, or “outer-loop” in the claims. Applicant’s arguments that the cited art do not teach what is in the specification is not relevant. Therefore, the rejection is proper an maintained.
Applicant argues “Zadeh is completely silent on "updating the first parameters
of the final layer based on a first training objective comparing the first time series data and the first outputs of the neural network while keeping the second parameters of the other layers frozen; generating, by the neural network parametrized with updated first parameters and the second parameters that have been frozen, second outputs based on an input of normalized time coordinates from the first horizon time window" of the original claim 1.” (Remarks, page 13, paragraph 1, line 1.) However, Examiner explained his interpretation that during backpropagation the prior layers are frozen while the latest layers are updated. Furthermore, claim 1 has been substantively amended. Because of this, the argument is moot in view of new grounds of rejection necessitated by amendment. See detailed rejection.
Applicant argues that new “Claims 21 and 22 should be found allowable because of their dependencies from independent claim 1, as well as their additional distinguishing features.” (Remarks, page 13, paragraph 4, line 1.) However, claim 1 remains rejected. Claims 21 and 22 are rejected as well.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claim(s) 1 - 18 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1
According to the first part of the analysis, in the instant case, claims 1 –18, and 21-22 are directed to a process, machine, manufacture, or composition of matter. As each of the claims falls within one of the four statutory categories (i.e., process, machine, manufacture, or composition of matter).
Independent Claim 1:
Step 2A Prong 1:
A method of training a time series data forecasting model, the method comprising:
generating, (This step recites generating an output based on an input of time coordinates which is practically implementable in the human mind with the aid of a pen and paper and is understood to be a recitation of a mental process. Generating an output by a general input based on time coordinates is something that can be performed on a piece of paper. )
training the neural network by updating the first parameters of the final layer based on a first training objective comparing the first time series data and the first outputs of the neural network while keeping the second parameters of the other layers frozen; (This step recites training a neural network by updating parameters of a layer by comparing data which is practically implementable in the human mind with the aid of a pen and paper and is understood to be a recitation of a mental process. Updating the parameters, regardless of the purpose can be understood to just be updating parameters stored on a piece of paper and forming a judgement if the data should be updated.)
generating, (This step recites generating an output based on an input of time coordinates which is practically implementable in the human mind with the aid of a pen and paper and is understood to be a recitation of a mental process. Generating an output by a general input based on time coordinates is something that can be performed on a piece of paper.)
and training a neural network by updating the second parameters based on a second training objective comparing the second time series data and second outputs of the neural network subject to the updated first parameters of the final layer. (This step recites training a neural network by updating parameters of a layer by comparing data which is practically implementable in the human mind with the aid of a pen and paper and is understood to be a recitation of a mental process. Updating the parameters, regardless of the purpose can be understood to just be updating parameters stored on a piece of paper and forming a judgement if the data should be updated.)
predicting, by the neural network, a set of future time series data over a future time period based on known time series data in the past. (Predicting a set of future time series data based on known time series data in the past is a mental process, something that can be performed by a human mind, or with the aid or pen and paper. The limitation “by the neural network” is mere instructions to apply.)
If a claim limitation under its broadest reasonable interpretation, covers performance of the limitation that can be performed in the human mind and/or using pen and paper as a physical aid, then it falls within the “Mental Processes” grouping of abstract ideas. Accordingly, the claim recites an abstract idea.
Step 2A Prong 2:
The judicial exception is not integrated into a practical application.
receiving a time-series data sequence including first time series data over a first lookback time window and second time series data over a first horizon time window following the lookback time window in time; (Receiving data is rejected under insignificant extra-solution activity. See MPEP 2106.05(g).)
by a neural network parametrized by first parameters of a final layer and second parameters of other layers (Parametrizing data and using neural networks to generate data is understood to be an mere instructions to apply an exception. See MPEP 2106.05(f).)
by the neural network parametrized with updated first parameters and the second parameters that have been frozen (Parametrizing data and using neural networks to generate data is understood to be an mere instructions to apply an exception. See MPEP 2106.05(f).)
Step 2B:
receiving a time-series data sequence including first time series data over a first lookback time window and second time series data over a first horizon time window following the lookback time window in time; (Receiving data is well-understood, routine, and conventional activities as supported under Berkheimer Evidence “Receiving or transmitting data over a network”. See MPEP 2106.05(d)(ll)(i).)
by a neural network parametrized by first parameters of a final layer and second parameters of other layers (Parametrizing data and using neural networks to generate data is understood to be an mere instructions to apply an exception. See MPEP 2106.05(f)(3).)
by the neural network parametrized with updated first parameters and the second parameters that have been frozen (Parametrizing data and using neural networks to generate data is understood to be an mere instructions to apply an exception. See MPEP 2106.05(f)(3).)
The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception because when considered separately and in combination, they do not add significantly more (also known as an “inventive concept”) to the exception.
Independent claims 8 and 15 recite the same relevant limitations and a similar analysis applies. Claim 8 recite the additional elements of “A system for training a time series data forecasting model” – components recited at a high level are construed as generic components used to implement the abstract idea. See MPEP 2106.05(f)(2); and “one or more hardware processors that read and execute the plurality of processor- executable instructions from the memory to perform operations” – components recited at a high level are construed as generic components used to implement the abstract idea. See MPEP 2106.05(f)(2). Independent claim 15 recites the additional elements of “A non-transitory, machine-readable medium comprising a plurality of instructions” – components recited at a high level are construed as generic components used to implement the abstract idea. See MPEP 2106.05(f)(2).
The additional elements do not integrate the abstract idea into a practical application. Nor do they amount to significantly more. Therefore, the independent claims are not patent eligible.
Dependent Claims 2 – 7 are also ineligible for the same reasons given with respect to claim 1. The dependent claims describe further mental processes and/or do not include additional active functional limitations/steps:
Claim 2:
Step 2A, Prong 1:
The method of claim 1, further comprising:
generating, (This step recites generating an output based on an input of time coordinates which is practically implementable in the human mind with the aid of a pen and paper and is understood to be a recitation of a mental process. Generating an output by a general input based on time coordinates is something that can be performed on a piece of paper.)
and updating the first parameters of the final layer based on the first training objective comparing third time series data over the second lookback time window and the third outputs of the neural network; while keeping the second parameters of the other layers frozen; (This step recites updating parameters by comparing data which is practically implementable in the human mind with the aid of a pen and paper and is understood to be a recitation of a mental process. Updating the parameters Is generically named here and can be understood to just be updating parameters stored on a piece of paper and forming a judgement if the data should be updated.)
Step 2A Prong 2 & Step 2B:
by the neural network (Using a neural network to generate data is understood to be mere instructions to apply an exception, in particular by using generic computing components. See MPEP 2106.05(f).)
Accordingly, under Step 2A Prong 2 and Step 2B, these additional elements do not integrate the abstract idea into practical application because they do not impose any meaningful limits on practicing the abstract idea, as discussed above in the rejection of claim 1.
Claim 3:
Step 2A, Prong 1:
The method of claim 2, further comprising:
generating, (This step recites generating an output based on an input of time coordinates which is practically implementable in the human mind with the aid of a pen and paper and is understood to be a recitation of a mental process. Generating an output by a general input based on time coordinates is something that can be performed on a piece of paper. )
and updating the second parameters based on the second training objective comparing fourth time series data over the second horizon time window and the fourth outputs of the neural network subject to the updated first parameters of the final layer. (This step recites updating parameters by comparing data which is practically implementable in the human mind with the aid of a pen and paper and is understood to be a recitation of a mental process. Updating the parameters Is generically named here and can be understood to just be updating parameters stored on a piece of paper and forming a judgement if the data should be updated.)
Step 2A Prong 2 & Step 2B:
by the neural network (Using a neural network to generate data is understood to be mere instructions to apply an exception, in particular by using generic computing components. See MPEP 2106.05(f).)
Accordingly, under Step 2A Prong 2 and Step 2B, these additional elements do not integrate the abstract idea into practical application because they do not impose any meaningful limits on practicing the abstract idea, as discussed above in the rejection of claim 1.
Claim 4:
Step 2A, Prong 1:
The method of claim 1, wherein the first training objective is computed by summing a cross entropy between the first time series data and the first outputs of the neural network over the first lookback time window. (This step generically recites computing a sum of cross entropy between two sets of data which is understood to be the abstract idea of a mathematical concept.)
If a claim limitation under its broadest reasonable interpretation, covers performance of the limitation that covers mathematical relationships, formulas, equations, or calculations then it falls under a recitation of a mathematical concept. Accordingly, the claim recites an abstract idea.
Step 2A Prong 2 & Step 2B:
Under Step 2A Prong 2, This judicial exception is not integrated into a practical application.
Under Step 2B, The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Claim 5:
Step 2A, Prong 1:
The method of claim 1, wherein the second training objective is computed by summing a cross entropy between the second time series data and the second outputs of the neural network over the first horizon time window. (This step generically recites computing a sum of cross entropy between two sets of data which is understood to be the abstract idea of a mathematical concept.)
If a claim limitation under its broadest reasonable interpretation, covers performance of the limitation that covers mathematical relationships, formulas, equations, or calculations then it falls under a recitation of a mathematical concept. Accordingly, the claim recites an abstract idea.
Step 2A Prong 2 & Step 2B:
Under Step 2A Prong 2, This judicial exception is not integrated into a practical application.
Under Step 2B, The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Claim 6:
Step 2A, Prong 1:
The method of claim 1, wherein the input of normalized time coordinates from the first lookback time window are modified by one or more sinusoid functions. (This step recites modifying an set of data by use of a sinusoid function which is practically implementable in the human mind with the aid of a pen and paper and is understood to be a recitation of a mental process. The claim generically recites that the sinusoid function is just used but does not make mention of how it is used, which means that it could that under broadest reasonable interpretation it could just be a function observed and then making a judgement call to modify the input.)
Step 2A Prong 2 & Step 2B:
Under Step 2A Prong 2, This judicial exception is not integrated into a practical application.
Under Step 2B, The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Claim 7:
Step 2A, Prong 1:
The method of claim 1, wherein the input of normalized time coordinates from the first lookback time window are modified by a concatenation of sinusoid functions. (This step recites modifying an set of data by use of a combination (concatenation) of sinusoid functions which is practically implementable in the human mind with the aid of a pen and paper and is understood to be a recitation of a mental process. The claim generically recites that the sinusoid function is just used but does not make mention of how it is used, which means that it could that under broadest reasonable interpretation it could just be a function observed and then making a judgement call to modify the input.)
Step 2A Prong 2 & Step 2B:
Under Step 2A Prong 2, This judicial exception is not integrated into a practical application.
Under Step 2B, The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
The additional elements do not integrate the abstract idea into a practical application. Nor do they amount to significantly more. Therefore, claims 2-7 are not patent eligible.
Claims 9-14 are system claims that correspond to method claims 2-7, respectively. Otherwise, they are not patentably distinct. A similar analysis applies. Therefore, claims 9-14 are not patent eligible.
Claims 16-18 are non-transitory, machine-readable medium claims that correspond to method claims 2-4, respectively. Otherwise, they are not patentably distinct. A similar analysis applies. Therefore, claims 16-18 are not patent eligible.
Therefore, claims 1-18 are not patent eligible.
Claims 21 and 22 are not rejected under U.S.C. § 101 as they cannot practically be performed by a human mind.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1-18, and 21-22 are rejected under 35 U.S.C. 103 as being unpatentable over Bianchi, et al (An overview and comparative analysis of Recurrent Neural Networks for Short Term Load Forecasting, herein Bianchi); Zadeh, et al (U.S. Publication US20180204111A1, System and Method for Extremely Efficient Image and Pattern Recognition and Artificial Intelligence Platform, herein Zadeh); and Saberi, et al (Deep Bayesian Recommendation Systems, herein Saberi).
Regarding Claim 1, Bianchi teaches:
A method of training a time series data forecasting model, the method comprising:
receiving a time-series data sequence including first time series data over a first lookback time window and second time series data over a first horizon time window following the lookback time window in time; (Bianchi teaches Figure 2 (below) on Page 5 which depicts the structure of an RNN- which also describes the variations of a received (general) input (X) including an XK-1 and a XK input which are separate intervals (windows) (the first window Xk-1 represents the first lookback time window as it is data collected previously before the input into the machine and correlating data and the second window and correlating data XK is the first horizon time window as its data which follows the lookback time window in time and represents a future (horizon) state of the data.). Furthermore, the collection of inputs x is described as time-series data: “The time series x has length T and it can contain real values, discrete values, one-hot vectors, and so on.” [Page 10]. Therefore, the model receives as an input the time-series data that has the two separate windows.)
Examiner Note: In regards to the limitations, the lookback time window and the horizon time window are understood to be linked, in the sense that the first lookback time window is a time window before the horizon time window (i.e. the time before the future (horizon) result). Bianchi teaches receiving information X, which would all be captured data from the past, with Xk just representing the data being analyzed at the certain time by the model.
generating, by a neural network parametrized by first parameters of a final layer and second parameters of other layers, first outputs based on an input of normalized time coordinates from the first lookback time window; (Bianchi teaches Figure 2 (below) which describes a final layer of weights which are combined with the input of X (which include normalized time coordinates from a first lookback time window) to generate an output Y. Weights are understood to combine to create an output as a function of the time input thus they are understood to be parametrized by the weight parameters of the layers. Furthermore, Bianchi teaches on page 20: “Beside differentiation, a common practice in STLF is to apply some form of normalization to the data” which teaches normalizing the time series data. Time coordinates is understood to be just the specific points of data within the lookback time window, therefore by applying normalization to the lookback time window, the time coordinates are also being normalized. Page 23 of Bianchi also teaches normalizing the time series data. Finally, Bianchi teaches an output of the system when set to training mode being the calculated Vk score [Page 5]. )
training the neural network, by updating the first parameters of the final layer based on a first training objective (Bianchi teaches applying Back Propagation Through Time (BPTT) on Page 5. Back Propagation is a method in the art that updates layers of a model starting from the back of the model and moving towards the start of the model. Whenever back propagation is applied, the first step (of Backpropagation) is updating the final layer while keeping the values of the previous layers the same (frozen). Bianchi teaches training the neural network by updating using BPTT based on a training objective, the objective being: “The gradient descent procedure consists in repeating two basic steps until convergence is reached” [Page 6] in which the gradient descent is the process of which backpropagation is applied.)
generating, by the neural network parametrized with updated first parameters and the second parameters that have been frozen, second outputs based on an input of normalized time coordinates from the first horizon time window; (It is understood that Bianchi teaches the steps above in regards to each epoch: “Note that with Wk we refer to all network parameters, while the index k identifies their values at epoch k, as they are updated during the optimization procedure.” [Page 6], therefore it is for each time window including the first horizon time window. It is understood that Bianchi first updates the weights based on the k (first lookback time window) and corresponding data Xk-1 then updates the weights by using the first horizon time window (k) and corresponding time series data Xk. The step of the creation of the Vk-1 score [Page 5] is one of the outputs of the neural network that is parameterized with the updated parameters based on the input horizon time window and time coordinates.)
Examiner Note: As the claims are currently recited, the backpropagation step still has created a neural network in which the first parameters were updated and while they were the second parameters that have been frozen while the first parameter update occurred.
and training the neural network by updating the second parameters based on a second training objective (Bianchi teaches applying Back Propagation Through Time (BPTT) on Page 5. At this step, the new generated output is being passed with Xk as the input used to create that output and the weights of the second parameters being updated by back propagation, in a system that is similar to the one given for the original updating step given above. As taught by the BPTT step, the weights are updated for each time the inputs are processed using the weights of the previous update. See Bianchi page 2, “Its output, at each time step, depends on previous inputs and past computations. This allows the network to develop a memory of previous events, which is implicitly encoded in its hidden state variables” in which the hidden state variables are the hidden weights, which include the updated first parameters of the final layer. As described above backpropagation is training a neural network.)
PNG
media_image1.png
634
1181
media_image1.png
Greyscale
predicting, by the trained neural network, a set of future time series data over a future time period based on known time series data in the past (Bianchi, page 10, paragraph 3, line 1 “In this section, we present three different RNN architectures trainable through the BPPT procedure, which we employ to predict real-valued time series. First, in Sec. 3.1 we present the most basic version of RNN, called Elman RNN. In Sec. 3.2 and 3.3 we discuss two gated architectures, which are LSTM and GRU. For each RNN model, we provide a quick overview of the main applications in time series forecasting and we discuss its principal features.” In other words, predict is predict, trainable is trained, RNN is neural network, and predict real-valued time series is predict a set of future time series data over a future time period.)
Thus far, Bianchi does not explicitly disclose:
comparing the first time series data and the first outputs of the neural network
Bianchi does teach comparing something that is created by the first time series data, however it is not directly the time series data. Bianchi, for example teaches: “In this first-order method, the weights Wk are updated according to a linear combination of the current gradient ∇Lk(Wk) and the previous update Vk−1, which is scaled by a hyperparameter µ: Vk = µVk−1 − η∇Lk(Wk), Wk+1 = Wk + Vk. (4)” [Page 7] which involve taking the output (current gradient) and the previous update Vk-1 which was generated by using the time series data and formulating an equation that takes both into consideration (which is understood as comparing the values) The Vk score being created by the inputs of the first time series data is supported by equation 1 on page 6 and equation 5 on page 8 in which Lk is the term that directly uses the series data.
However, Zadeh teaches a comparison directly as required.
comparing the first time series data and the first outputs of the neural network (Zadeh teaches: “In one embodiment, the objective function (error function) to optimize in back propagation, is the cross entropy error, Es, between the data (e.g., image pixel intensity in V layer) and the reconstruction (e.g., the corresponding pixel intensities in V′ output), for a given sample” [1670] which teaches updating the model by backpropagation by applying a comparison between an output and the inputted data (See Paragraph 1693 in which the V layer is supplied input images, which are the original input data.) Backpropagation is a means for updating the first parameters while keeping the other layers frozen as explained above. Applying Cross-entropy calculations is understood to be comparing the data, as the equation given after 1670 shows that both time series data and first outputs are elements that directly control the outputted value and thus interact with each other to form a comparison.)
Before the effective filing date of the invention it would have been obvious to one of ordinary skill in the art to modify the gradient calculation process used in the backpropagation of Bianchi with the cross entropy objective function process as taught by Zadeh with the comparison of the inputted data and the output for the improvement of being able to reconstruct the input data accurately. This improvement is taught by Zadeh: “In one embodiment, the feature detector is fine tuned, e.g., via a deep autoencoder and back propagation, to be able to reconstruct images accurately.” [2510]
Examiner note: This modification propagates for every backpropagation step of Bianchi, as the back propagation method as taught of Bianchi is performed for the process when using the lookback time data or the horizon time data. Bianchi also teaches: “As an RNN processes sequential information, it performs the same operations on every element of the input sequence.” [Page 2] which teach the operations being performed for every inputted set of data.
Thus far, the combination of Bianchi and Zadeh does not explicitly teach
the second parameters of the other layers frozen.
However, Saberi teaches the second parameters of the other layers frozen (Saberi, page 3, paragraph 6, line 5 “This is described in Algorithm 1. Approximate knowledge of the posterior is sufficient for TS, which is one of our subroutines; to do UCB, we first partition the parameters as Θt = (Θ<L, ΘMAPL ). We view θ< L as “ fixed” frozen parameters, and ΘMAPL as the maximum a posteriori (MAP) estimate for the last layer as obtained by the training. So, we can compute fΘ<L(xt,a) for each action.” In other words, for Θ<L frozen parameters is second layers of the other layers frozen.)
Both Saberi and the combination of Bianchi and Zadeh are directed to neural networks and forecasting, among other things. The combination of Bianchi and Zadeh teach a method of training a time series data forecasting model, but does not explicitly teach training for an objective while keeping the second parameters of other layers frozen. Saberi teaches training for an objective while keeping the second parameters of other layers frozen. It would be obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teaching of Saberi into the combination of Bianchi and Zadeh. This would result in a method of training a time series data forecasting model, and training for an objective while keeping the second parameters of other layers frozen.
One of ordinary skill in the art would be motivated to do this in order to improve recommendations and approximations of reward uncertainties (Saberi, abstract, line 3 “The presence of deep neural networks can lead to better recommendations, but with the caveat of
more difficult and costly uncertainty estimates. Importantly, modeling the risk in a recommendation helps to better navigate the explore/exploit tradeoff, i.e., to decide whether to recommend similar or new, novel content based on available context. In this paper, we explore using approximate Bayesian Neural Networks to yield better approximations of reward uncertainties, and implement our method with both Thompson Sampling and Upper Confidence Bound policies.”)
Regarding Claim 2, Bianchi in view of Zadeh, in further view of Saberi teaches all the limitations of Claim 1, and further teaches: The method of claim 1, further comprising:
generating, by the neural network, third outputs based on an input of normalized time coordinates from a second lookback time window of the time-series data sequence; (Bianchi teaches Figure 2, which shows the inputted data Xk+1 which represents the second lookback time window of the time-series data with a corresponding output Yk+1. As taught by Claim 1, the data is normalized before inputted and as taught on page 20 of Bianchi.)
and updating the first parameters of the final layer based on the first training objective comparing third time series data over the second lookback time window and the third outputs of the neural network; while keeping the second parameters of the other layers frozen; (Bianchi teaches applying Back Propagation Through Time (BPTT) on Page 5. At this step, the new generated output is being passed with Xk+1 as the input used to create that output and the weights of the first parameters being updated by back propagation, in a system that is similar to the one given for the original updating step given above in Claim 1.
Regarding it specifically updating using the lookback time window and data, Zadeh teaches preforming backpropagation in paragraph 1670 as explained in Claim 1 that uses the collected data in the V layer and the output.)
Reasons for obviousness has been recited in Claim 1 above.
Regarding Claim 3, Bianchi in view of Zadeh, in further view of Saberi teaches all the limitations of Claim 1, and further teaches: The method of claim 1, further comprising:
generating, by the neural network, fourth outputs based on an input of normalized time coordinates from a second horizon time window; (Figure 2 only directly represents 3 sets of data, however as understood in the art and as described by the ellipses of the data after the xk+1 step, this is meant to continue on for any set of input data, including a possible XK+2 step which would represent a second horizon time window which is a horizon of the XK+1 data. As taught by Claim 1, the data is normalized before inputted into the system.)
and updating the second parameters based on the second training objective comparing fourth time series data over the second horizon time window and the fourth outputs of the neural network subject to the updated first parameters of the final layer. (Bianchi teaches applying Back Propagation Through Time (BPTT) on Page 5. At this step, the new generated output is being passed with Xk+2 as the input used to create that output and the weights of the second parameters being updated by back propagation, in a system that is similar to the one given for the original updating step given above. As taught by the BPTT step, the weights are updated for each time the inputs are processed using the weights of the previous update. See Bianchi page 2, “Its output, at each time step, depends on previous inputs and past computations. This allows the network to develop a memory of previous events, which is implicitly encoded in its hidden state variables” in which the hidden state variables are the hidden weights, which include the updated first parameters of the final layer. Regarding the training objective and the backpropagation step specifically updating using the horizon time window and data, Zadeh teaches preforming backpropagation in paragraph 1670 as explained in Claim 1.)
Reasons for obviousness has been recited in Claim 1 above.
Regarding Claim 4, Bianchi in view of Zadeh, in further view of Saberi teaches all the limitations of Claim 1, and further teaches: The method of claim 1, wherein the first training objective is computed by summing a cross entropy between the first time series data and the first outputs of the neural network over the first lookback time window. (As explained in Claim 1, the back propagation step of Bianchi teaches the inputted data of a first lookback time window and the current modification with Zadeh teaches a specific training objective of summing a cross entropy between an input data and output of a neural network: “In one embodiment, the objective function (error function) to optimize in back propagation, is the cross entropy error, Es, between the data (e.g., image pixel intensity in V layer) and the reconstruction (e.g., the corresponding pixel intensities in V′ Output), for a given sample.” [1670] as well as an equation showing the summation of these scores to generate Es which is represented by the equation provided directly below the quotation provided which involves calculating a summation of the cross-entropy calculations.)
Reasons for obviousness has been recited in Claim 1 above.
Regarding Claim 5, Bianchi in view of Zadeh, in further view of Saberi teaches all the limitations of Claim 1, and further teaches: The method of claim 1, wherein the second training objective is computed by summing a cross entropy between the second time series data and the second outputs of the neural network over the first horizon time window. (As explained in Claim 1, the back propagation step of Bianchi teaches the inputted data of a first horizon time window and the current modification with Zadeh teaches a specific training objective of summing a cross entropy between an input data and output of a neural network: “In one embodiment, the objective function (error function) to optimize in back propagation, is the cross entropy error, Es, between the data (e.g., image pixel intensity in V layer) and the reconstruction (e.g., the corresponding pixel intensities in V′ Output), for a given sample.” [1670] as well as an equation showing the summation of these scores to generate Es which is represented by the equation provided directly below the quotation provided which involves calculating a summation of the cross-entropy calculations.)
Reasons for obviousness has been recited in Claim 1 above.
Regarding Claim 6, Bianchi in view of Zadeh, in further view of Saberi teaches all the limitations of Claim 1, and Bianchi further teaches: The method of claim 1, wherein the input of normalized time coordinates from the first lookback time window are modified by one or more sinusoid functions. (Bianchi teaches generating a synthetic time series by applying a multiple superimposed oscillator on the original time series, as taught on Page 18 in which a time interval (time window) tf is inputted into the function for y(t) to generate the synthetic time series data modified by one or more sinusoid functions (see equation 30 with the modification using a sinusoid function). Furthermore, Bianchi teaches on page 20: “Beside differentiation, a common practice in STLF is to apply some form of normalization to the data” which teaches normalizing the time series data to be used in the system.)
PNG
media_image2.png
432
1164
media_image2.png
Greyscale
Regarding Claim 7, Bianchi in view of Zadeh, in further view of Saberi teaches all the limitations of Claim 1, and Bianchi further teaches: The method of claim 1, wherein the input of normalized time coordinates from the first lookback time window are modified by a concatenation of sinusoid functions. (Bianchi teaches a further step of generating a synthetic time series by applying a multiple superimposed oscillator on the original time series, as taught on Page 18 in which a time interval (time window) tf is inputted into the function for y(t) to generate the synthetic time series data modified by a concatenation of sinusoid functions (see equation 30 with the modification using a sinusoid function). Furthermore, Bianchi teaches on page 20: “Beside differentiation, a common practice in STLF is to apply some form of normalization to the data” which teaches normalizing the time series data to be used in the system.)
Claims 8-14 are system claims corresponding to method claims 1-7, respectively. Otherwise, they are not patentably distinct. Claims 8-14 recite the additional elements of “A system… and a plurality of processor executable instructions…comprising one or more hardware processors that read and execute the plurality of processor- executable instructions from the memory to perform operations” One of ordinary skill in the art would understand the primary reference as teaching a system with a processor, memory, and instructions as being required for the execution of a computer implemented method. “[I]n considering the disclosure of a reference, it is proper to take into account not only specific teachings of the reference but also the inferences which one skilled in the art would reasonably be expected to draw therefrom.” MPEP § 2144.01. Therefore, claims 8-14 are rejected for the same reasons as claims 1-7, respectively.
Claims 15-18 are non-transitory machine readable medium claims corresponding to method claims 1-4, respectively. Otherwise, they are not patentably distinct. Claims 15-18 recite the additional elements of a non-transitory machine readable medium. One of ordinary skill in the art would understand the primary reference as teaching a non-transitory machine readable medium as being required for the execution of a computer implemented method. “[I]n considering the disclosure of a reference, it is proper to take into account not only specific teachings of the reference but also the inferences which one skilled in the art would reasonably be expected to draw therefrom.” MPEP § 2144.01. Therefore, claims 15-18 are rejected for the same reasons as claims 1-4, respectively
Regarding Claim 21, Bianchi in view of Zadeh, in further view of Saberi teaches all the limitations of Claim 1, and Bianchi further teaches: The method of claim 1, wherein the
keeping of the second parameters of the other layers frozen includes keeping the second parameters of the other layers during the generating of the second outputs (Saberi, Algorithm 1, and page 3, paragraph 5 “We first model rewards by a fully connected neural network, the same approach as [14]. Then, after training the network to yield parameters Θt, we use a Laplace approximation to model the posterior distribution of the neural network in lieu of exact posterior inference:
PNG
media_image3.png
28
332
media_image3.png
Greyscale
However, even this appoximation is intractible for deep FCNNs, since even small FCNNs will have (at least) thousands of parameters, and sampling from Gaussians (with dense covariance matrices) inthat regime can be very costly. A typical approximation that is made for numerical effiency is just adopting a “last-layer” Laplace approximation [8], where the posterior distribution is only over the last-layer parameters of the FCNN (and all others are frozen at inference time). This is described in Algorithm 1.”
PNG
media_image4.png
544
922
media_image4.png
Greyscale
In other words, frozen parameters is frozen parameters and adopting a “last layer” while all other parameters are frozen is keeping the second parameters of the other layers frozen during generating outputs.)
Regarding claim 22, Bianchi in view of Zadeh, in further view of Saberi teaches all the limitations of Claim 1, and Bianchi further teaches: The method of claim 1, the updating of the second parameters based on a second training objective includes keeping the updated first parameters frozen prior to the updated of the second parameters (Saberi, Algorithm 1, and page 3, paragraph 5. See above mapping. In other words, adopting a “last layer” is updating the second parameters for the second objective, and all other layers frozen is keeping the updated first parameter frozen prior to the updated second parameters.)
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 BART RYLANDER whose telephone number is (571)272-8359. The examiner can normally be reached Monday - Thursday 8:00 to 5:30.
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, Miranda Huang can be reached at 571-270-7092. 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.
/B.I.R./Examiner, Art Unit 2124
/VINCENT GONZALES/Primary Examiner, Art Unit 2124
Prosecution Insights