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 Arguments
Applicant's arguments filed 07/07/2026 have been fully considered but they are not persuasive.
Regarding the 101 rejections, on pages 11-12 of “Remarks” applicant contends that the amended claim 1 provides a practical application under Step 2A Prong 2. The examiner respectfully disagrees. Specifically, applicant argues that the amended limitation of and using, by the at least one processor, the inference output as a task result for a given task associated with the inference data incorporates any alleged judicial exceptions into a practical application by using the inference output as a task result for a given task associated with the inference data. Under the broadest reasonable interpretation, the amended limitation recite mere instructions to apply an exception, because they recite no more than an idea of a solution or outcome, like using a model’s output for a task, which represents merely adding the words “apply it”, or an equivalent, which are not indicative of an inventive concept. See MPEP 2106.05(f)(1): “The recitation of claim limitations that attempt to cover any solution to an identified problem with no restriction on how the result is accomplished and no description of the mechanism for accomplishing the result, does not integrate a judicial exception into a practical application or provide significantly more because this type of recitation is equivalent to the words "apply it". See Electric Power Group, LLC v. Alstom, S.A., 830 F.3d 1350, 1356, 119 USPQ2d 1739, 1743-44 (Fed. Cir. 2016); Intellectual Ventures I v. Symantec, 838 F.3d 1307, 1327, 120 USPQ2d 1353, 1366 (Fed. Cir. 2016); Internet Patents Corp. v. Active Network, Inc., 790 F.3d 1343, 1348, 115 USPQ2d 1414, 1417 (Fed. Cir. 2015). In contrast, claiming a particular solution to a problem or a particular way to achieve a desired outcome may integrate the judicial exception into a practical application or provide significantly more. See Electric Power, 830 F.3d at 1356, 119 USPQ2d at 1743.”. Therefore, applicant’s arguments regarding the 101 rejections are not persuasive.
Regarding the 103 rejections, applicant's arguments filed with respect to the prior art rejections have been fully considered but they are moot. Applicant has amended the claims to recite new combinations of limitations. Applicant's arguments are directed at the amendment. Please see below for new grounds of rejection, necessitated by Amendment.
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.
Claims 1-18 are rejected under 35 U.S.C 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1, in step 1 of the 101 analysis set forth in MPEP 2106, the claim recites A method of using a Neural Network (NN),. The claim recites a method. A method is one of the four statutory categories of invention.
In Step 2A, Prong 1 of the 101 analysis set forth in MPEP 2106, the examiner has determined that the following limitations recite a process that, under broadest reasonable interpretation, covers a mental process or mathematical concept but for the recitation of generic computer components:
during a first training iteration of the NN: determining a first continuous sequence of intermediate layers from the plurality of intermediate layers of the NN, wherein determining the first continuous sequence comprises randomly sampling the first continuous sequence from a predefined distribution over a sample space of continuous sequences of intermediate layers of the NN, the input layer, the first continuous sequence of intermediate layers, and the output layer forming a first sub-network of the NN; (i.e., the broadest reasonable interpretation includes a step of observation, evaluation, and judgement and could be performed mentally or with pen and paper like randomly selecting layers from a group to form a first subnetwork, which is either a mental process of observation/evaluation/judgement (MPEP 2106)).
training the first sub-network based on training data, the training of the first sub-network comprising calculating a first plurality of gradients for parameters associated with the first sub-network; (i.e., the broadest reasonable interpretation includes mathematical calculation like calculating a gradient, a mathematical calculation is considered a mathematical concept (MPEP 2106)).
during a second training iteration of the NN: determining a second continuous sequence of intermediate layers from the plurality of intermediate layers of the NN, wherein determining the second continuous sequence comprises randomly sampling the second continuous sequence from the predefined distribution over the sample space of continuous sequences of intermediate layers of the NN, the input layer, the second continuous sequence of intermediate layers, and the output layer forming a second sub-network of the NN, the second continuous sequence of intermediate layers being different from the first continuous sequence of intermediate layers, the second continuous sequence of intermediate layers at least partially overlapping the first continuous sequence of intermediate layers; (i.e., the broadest reasonable interpretation includes a step of observation, evaluation, and judgement and could be performed mentally or with pen and paper like randomly selecting second layers from a group to form a second subnetwork that overlaps with the first subnetwork, which is either a mental process of observation/evaluation/judgement (MPEP 2106)).
training the second sub-network based on the training data, the training of the second sub-network comprising calculating a second plurality of gradients for parameters associated with the second sub-network; (i.e., the broadest reasonable interpretation includes mathematical calculation like calculating a gradient, a mathematical calculation is considered a mathematical concept (MPEP 2106)).
after the first training iteration and the second training iteration, updating parameters of the NN based on accumulated gradients comprising the first plurality of gradients and the second plurality of gradients; (i.e., the broadest reasonable interpretation includes mathematical calculation like summing gradients, a mathematical calculation is considered a mathematical concept (MPEP 2106)).
during an inference iteration of the NN: selecting a target sub-network amongst the first sub-network and the second sub-network; (i.e., the broadest reasonable interpretation includes a step of observation, evaluation, and judgement and could be performed mentally or with pen and paper like selecting the subnetwork that is the fastest, which is either a mental process of observation/evaluation/judgement (MPEP 2106)).
If the claim limitations, under their broadest reasonable interpretation, covers activities classified under Mental processes: concepts performed in the human mind (including observation, evaluation, judgement, or opinion) (see MPEP 2106.04(a)(2), subsection (III)) or Mathematical concepts: mathematical relationships, mathematical formulas or equations, or mathematical calculations (see MPEP 2106.04(a)(2), subsection (I)). Accordingly, the claim recites an abstract idea.
In Step 2A, Prong 2 of the 101 analysis, set forth in MPEP 2106, the examiner has determined that the following additional elements do not integrate this judicial exception into a practical application:
the NN comprising an input layer, an output layer, and a plurality of intermediate layers, the method executable by at least one processor and comprising: (i.e., the generic computer components recited in this limitation merely add the words “apply it”, or an equivalent, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f))).
generating an inference output by employing only the target sub-network of the NN on inference data for reducing computational resources of the at least one processor for generating the inference output, (i.e., the broadest reasonable interpretation of outputting a data instance is mere data outputting, which is an insignificant extra solution activity (MPEP 2106.05(g))).
and using, by the at least one processor, the inference output as a task result for a given task associated with the inference data. (i.e., the generic computer components recited in this limitation merely add the words “apply it”, or an equivalent, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f))).
Since the claim does not contain any other additional elements, that amount to integration into a practical application, the claim is directed to an abstract idea.
In Step 2B of the 101 analysis set forth in the 2019 PEG, the examiner has determined that the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception:
Regarding limitation (VIII), under the broadest reasonable interpretation, recite steps of mere data gathering/outputting, which has been recognized by the courts as being well-understood, routine, and conventional functions. Specifically, the courts have recognized computer functions directed to mere data gathering/outputting as well-understood, routine, and conventional functions when they are claimed in a merely generic manner or as insignificant extra-solution activity when considering evidence in view of Berkheimer v. HP, Inc., 881 F.3d 1360, 1368, 125 USPQ2d 1649, 1654 (Fed. Cir. 2018), see USPTO Berkheimer Memorandum (April 2018)).
Examiner uses Berkheimer: Option 2, a citation to one or more of the court decisions discussed in MPEP 2106.05(d)(II) as noting well-understood, routine, and conventional nature of the additional elements:
Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016) (using a telephone for image transmission); OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363, 115 USPQ2d 1090, 1093 (Fed. Cir. 2015) (sending messages over a network); buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network). See MPEP 2106.05(d)(II).
Further, limitation (VII), under the broadest reasonable interpretation, merely recite steps that apply a generic computer component and a generic neural network as tools to perform judicial exceptions, which represents merely adding the words “apply it”, or an equivalent, which are not indicative of an inventive concept (MPEP 2106.05(f)). Similarly, limitation (IX), under the broadest reasonable interpretation, recite mere instructions to apply an exception, because they recite no more than an idea of a solution or outcome which represents merely adding the words “apply it”, or an equivalent, which are not indicative of an inventive concept (MPEP 2106.05(f)). Considering additional elements individually and in combination, and the claim as a whole, the additional elements do not provide significantly more than the abstract idea. Therefore, the claim is not patent eligible.
Regarding claim 2, it is dependent upon claim 1 and fails to resolve the deficiencies identified above by integrating the judicial exception into a practical application, or introducing significantly more than the judicial exception. For example, claim 2 recites during the first training iteration: determining a first depth index indicative of a first depth of the first continuous sequence of intermediate layers in the NN, and wherein the determining the first continuous sequence of intermediate layers includes: determining the continuous sequence of intermediate layers that is most adjacent to the input layer of the NN and which includes a total number of layers equal to the first depth index;. Under the broadest reasonable interpretation, the limitations recite determining a depth of the first subnetwork which is a step of observation, evaluation, and judgement which can be performed mentally or with pen and paper. Claim 2 also recites and wherein: during the second training iteration: determining a second depth index indicative of a second depth of the second continuous sequence of intermediate layers in the NN, the second depth index being different from the first depth index, and wherein the determining the second continuous sequence of intermediate layers includes: determining an other continuous sequence of intermediate layers that is most adjacent to the input layer of the NN and which includes a total number of layers equal to the second depth index. Under the broadest reasonable interpretation, the limitations recite determining a depth of the second subnetwork which is a step of observation, evaluation, and judgement which can be performed mentally or with pen and paper. The steps of observation, evaluation, and judgement are mental processes. Therefore, claim 2 does not solve the deficiencies of claim 1.
Regarding claim 3, it is dependent upon claim 2 and fails to resolve the deficiencies identified above by integrating the judicial exception into a practical application, or introducing significantly more than the judicial exception. For example, claim 3 recites wherein randomly sampling the first continuous sequence comprises randomly determining the first depth index from an interval of depth indexes, and randomly sampling the second continuous sequence comprises randomly determining the second depth index from the interval of depth indexes, the interval of depth indexes having been pre-determined based on a depth of the NN. Under the broadest reasonable interpretation, the limitations recite randomly selecting depths from a range of depths which is a step of observation, evaluation, and judgement which can be performed mentally or with pen and paper. The steps of observation, evaluation, and judgement are mental processes. Therefore, claim 3 does not solve the deficiencies of claim 2.
Regarding claim 4, it is dependent upon claim 1 and fails to resolve the deficiencies identified above by integrating the judicial exception into a practical application, or introducing significantly more than the judicial exception. For example, claim 4 recites wherein the method further comprises: during the first training iteration: determining a first width index for the first continuous sequence of intermediate layers indicative of a first partial width of the first continuous sequence of intermediate layers to be trained during the first training iteration,. Under the broadest reasonable interpretation, the limitations recite determining a width of the first subnetwork which is a step of observation, evaluation, and judgement which can be performed mentally or with pen and paper. Claim 4 also recites and wherein the training the first sub-network includes: training only the first partial width of the first continuous sequence of intermediate layers based on the training data, Under the broadest reasonable interpretation, the limitations merely recite steps that apply generic training to a subset of parameters, which represents merely adding the words “apply it”, or an equivalent, which are not indicative of an inventive concept (MPEP 2106.05(f)). Claim 4 also recites and wherein: during the second training iteration: determining a second width index for the second continuous sequence of intermediate layers indicative of a second partial width of the second continuous sequence of intermediate layers to be trained during the second training iteration, the second width index being different from the first width index,. Under the broadest reasonable interpretation, the limitations recite determining a width of the second subnetwork which is a step of observation, evaluation, and judgement which can be performed mentally or with pen and paper. The steps of observation, evaluation, and judgement are mental processes. Claim 4 also recites the training the second sub-network includes: training only the second partial width of the second continuous sequence of intermediate layers based on the training data. Under the broadest reasonable interpretation, the limitations merely recite steps that apply generic training to a subset of parameters, which represents merely adding the words “apply it”, or an equivalent, which are not indicative of an inventive concept (MPEP 2106.05(f)). Therefore, claim 4 does not solve the deficiencies of claim 1.
Regarding claim 5, it is dependent upon claim 4 and fails to resolve the deficiencies identified above by integrating the judicial exception into a practical application, or introducing significantly more than the judicial exception. For example, claim 5 recites wherein the determining the first width index includes randomly determining the first width index from an interval of width indexes, and the determining the second width index includes randomly determining the second width index from the interval of width indexes, the interval of width indexes having been pre-determined based on a width of the plurality of intermediate layers of the NN. Under the broadest reasonable interpretation, the limitations recite randomly selecting widths from a range of widths which is a step of observation, evaluation, and judgement which can be performed mentally or with pen and paper. The steps of observation, evaluation, and judgement are mental processes. Therefore, claim 5 does not solve the deficiencies of claim 4.
Regarding claim 6, it is dependent upon claim 1 and fails to resolve the deficiencies identified above by integrating the judicial exception into a practical application, or introducing significantly more than the judicial exception. For example, claim 6 recites wherein the selecting the target sub-network comprises comparing at least one of:
(i) a first accuracy parameter of the first sub-network and a second accuracy parameter of the second sub-network,
(ii) a first latency parameter of the first sub-network and a second latency parameter of the second sub-network, and
(iii) a first importance parameter of the first sub-network and a second importance parameter of the second sub-network.
Under the broadest reasonable interpretation, the limitations recite selecting a subnetwork by comparing subnetworks with a criterion which is a step of observation, evaluation, and judgement which can be performed mentally or with pen and paper. The steps of observation, evaluation, and judgement are mental processes. Therefore, claim 6 does not solve the deficiencies of claim 1.
Regarding claim 7, it is dependent upon claim 1 and fails to resolve the deficiencies identified above by integrating the judicial exception into a practical application, or introducing significantly more than the judicial exception. For example, claim 7 recites wherein the plurality of intermediate layers is a plurality of architectural blocks of the NN, a given one of the plurality of architectural blocks including a sub-set of intermediate layers for generating an output of the given one of the plurality of architectural blocks. Under the broadest reasonable interpretation, the limitations recite grouping layers into blocks which is a step of observation, evaluation, and judgement which can be performed mentally or with pen and paper. The steps of observation, evaluation, and judgement are mental processes. Therefore, claim 7 does not solve the deficiencies of claim 1.
Regarding claim 8, it is dependent upon claim 7 and fails to resolve the deficiencies identified above by integrating the judicial exception into a practical application, or introducing significantly more than the judicial exception. For example, claim 8 recites wherein the plurality of architectural blocks include at least one of:
(i) a convolutional block with at least one convolutional layer;
(ii) a pooling block with at least one pooling layer;
(iii) a fully connected block with at least one fully-connected layer;
(iv) a residual block with at least one skip connection;
(v) a batch normalization block with at least one batch normalization layer;
(vi) a recurrent block with at least one recurrence mechanism;
(vii) an attention block with at least one self-attention mechanism; and
(viii) an activation block with at least one activation layer.
Under the broadest reasonable interpretation, the limitations merely recite steps that apply generic components of machine learning models, which represents merely adding the words “apply it”, or an equivalent, which are not indicative of an inventive concept (MPEP 2106.05(f)). Therefore, claim 8 does not solve the deficiencies of claim 7.
Regarding claim 9, it is dependent upon claim 1 and fails to resolve the deficiencies identified above by integrating the judicial exception into a practical application, or introducing significantly more than the judicial exception. For example, claim 9 recites wherein the output layer is at least two output layers, and wherein the output layer of the first sub-network is a first one from the at least two output layers, and the output layer of the second sub-network is a second one from the at least two output layers, the first and second one of the at least two output layers being different output layers. Under the broadest reasonable interpretation, the limitations recite determining an output layer for each subnetwork which is a step of observation, evaluation, and judgement which can be performed mentally or with pen and paper. The steps of observation, evaluation, and judgement are mental processes. Therefore, claim 9 does not solve the deficiencies of claim 1.
Regarding claim 10, in step 1 of the 101 analysis set forth in MPEP 2106, the claim recites A system for using a Neural Network (NN), the NN comprising an input layer, an output layer, and a plurality of intermediate layers, the system comprising at least a processor and a memory storing a plurality of executable instructions which, when executed by the at least a processor, cause the system to:. The claim recites system with hardware which is interpreted as a machine. A machine is one of the four statutory categories of invention. For the Step 2A/2B analyses, since claim 10 is similar to claim 1 it is rejected under the same rationales as claim 1.
The additional limitation below fails to resolve the deficiencies identified above by integrating the judicial exception into a practical application, or introducing significantly more than the judicial exception.
A system for using a Neural Network (NN), the NN comprising an input layer, an output layer, and a plurality of intermediate layers, the system comprising at least a processor and a memory storing a plurality of executable instructions which, when executed by the at least a processor, cause the system to:… (i.e., the generic computer components recited in this limitation merely add the words “apply it”, or an equivalent, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea (MPEP 2106.05(f))).
Considering additional elements individually and in combination, and the claim as a whole, the additional elements do not provide significantly more than the abstract idea. Therefore, the claim is not patent eligible.
Regarding claims 11-18, the claims are similar to claims 2-9 and rejected under the same rationales.
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 1-3, 6-12, and 15-18 are rejected under 35 U.S.C. 103 as being unpatentable over Teerapittayanon, et al., “BranchyNet: Fast Inference via Early Exiting from Deep Neural Networks” (“Teerapittayanon”) in view of Huang, et al., Non-Patent Literature “Deep Networks with Stochastic Depth” (“Huang”).
Regarding claim 1, Teerapittayanon discloses:
A method of using a Neural Network (NN), the NN comprising an input layer, an output layer, and a plurality of intermediate layers, (Teerapittayanon, pg. 2464 col. 2 and Figure 1, “Figure 1 shows how BranchyNet modifies a standard AlexNet by adding two branches with their respective exit points; Figure 1 shows an example layout of BranchyNet had how it has input, output, and hidden layers (i.e. A method of using a Neural Network (NN), the NN comprising an input layer, an output layer, and a plurality of intermediate layers,).”).
the method executable by at least one processor (Teerapittayanon, pg. 2467 col. 1, “We use a 3.0GHz CPU [the method executable by at least one processor] with 20 MB L3 Cache and NVIDIA GeForce GTX TITANX (Maxwell) 12GB GPU.”).
and comprising: during a first training iteration of the NN: determining a first continuous sequence of intermediate layers from the plurality of intermediate layers of the NN,…the input layer, the first continuous sequence of intermediate layers, and the output layer forming a first sub-network of the NN; (Teerapittayanon, pg. 2465-2466 and Figure 1, “BranchyNet modifies the standard deep network structure by adding exit branches (also called side branches or simply branches for brevity), at certain locations throughout the network. These early exit branches allow samples which can be accurately classified in early stages of the network to exit at that stage; in Figure 1, the layers that are part of the path to the first exit is interpreted as the first sub-network (i.e. and comprising: during a first training iteration of the NN: determining a first continuous sequence of intermediate layers from the plurality of intermediate layers of the NN,…the input layer, the first continuous sequence of intermediate layers, and the output layer forming a first sub-network of the NN;).”).
training the first sub-network based on training data, the training of the first sub-network comprising calculating a first plurality of gradients for parameters associated with the first sub-network; (Teerapittayanon, pg. 2466 col. 2, “The design goal of each exit branch is to minimize this loss function. To train the entire BranchyNet, we form a joint optimization problem as a weighted sum of the loss functions of each exit branch [training the first sub-network based on training data]”, and Teerapittayanon, pg. 2466 col. 2, “In the feedforward pass, the training data set is passed through the network, including both main and side branches, the output from the neural network at all exit points is recorded, and the error of the network is calculated. In backward propagation, the error is passed back through the network and the weights are updated using gradient descent [the training of the first sub-network comprising calculating a first plurality of gradients for parameters associated with the first sub-network;].”).
during a second training iteration of the NN: determining a second continuous sequence of intermediate layers from the plurality of intermediate layers of the NN,…the input layer, the second continuous sequence of intermediate layers, and the output layer forming a second sub-network of the NN, the second continuous sequence of intermediate layers being different from the first continuous sequence of intermediate layers, (Teerapittayanon, pg. 2465-2466 and Figure 1, “BranchyNet modifies the standard deep network structure by adding exit branches (also called side branches or simply branches for brevity), at certain locations throughout the network. These early exit branches allow samples which can be accurately classified in early stages of the network to exit at that stage; in Figure 1, the path to the second exit is interpreted as the second sub-network (i.e. during a second training iteration of the NN: determining a second continuous sequence of intermediate layers from the plurality of intermediate layers of the NN,…the input layer, the second continuous sequence of intermediate layers, and the output layer forming a second sub-network of the NN, the second continuous sequence of intermediate layers being different from the first continuous sequence of intermediate layers,).”).
the second continuous sequence of intermediate layers at least partially overlapping the first continuous sequence of intermediate layers; (Teerapittayanon, see Figure 1, Figure 1 shows that the second exit path utilizes the first 5x5 Conv layer of the base model which is also used by the first exit path (i.e. the second continuous sequence of intermediate layers at least partially overlapping the first continuous sequence of intermediate layers;)).
training the second sub-network based on the training data, the training of the second sub-network comprising calculating a second plurality of gradients for parameters associated with the second sub-network; (Teerapittayanon, pg. 2466 col. 2, “The design goal of each exit branch is to minimize this loss function. To train the entire BranchyNet, we form a joint optimization problem as a weighted sum of the loss functions of each exit branch [training the second sub-network based on the training data,] and Teerapittayanon, pg. 2466 col. 2, “In the feedforward pass, the training data set is passed through the network, including both main and side branches, the output from the neural network at all exit points is recorded, and the error of the network is calculated. In backward propagation, the error is passed back through the network and the weights are updated using gradient descent [the training of the second sub-network comprising calculating a second plurality of gradients for parameters associated with the second sub-network;].”).
during an inference iteration of the NN: selecting a target sub-network amongst the first sub-network and the second sub-network; (Teerapittayanon, pg. 2465 col. 1, “Once the network is trained [during an inference iteration of the NN:], BranchyNet utilizes the exit points to allow the samples to exit early, thus reducing the cost of inference; using an exit point is interpreted as selecting a subnetwork (i.e. selecting a target sub-network amongst the first sub-network and the second sub-network;).”).
after the first training iteration and the second training iteration, updating parameters of the NN based on accumulated gradients comprising the first plurality of gradients and the second plurality of gradients; (Teerapittayanon, pg. 2466 col. 2, “In the feedforward pass, the training data set is passed through the network, including both main and side branches [after the first training iteration and the second training iteration,], the output from the neural network at all exit points is recorded, and the error of the network is calculated. In backward propagation, the error is passed back through the network and the weights are updated using gradient descent [updating parameters of the NN based on accumulated gradients comprising the first plurality of gradients and the second plurality of gradients;].”).
generating an inference output by employing only the target sub-network of the NN on inference data for reducing computational resources of the at least one processor for generating the inference output, and using, by the at least one processor, the inference output as a task result for a given task associated with the inference data. (Teerapittayanon, pg. 2465 col. 1, “Once the network is trained, BranchyNet utilizes the exit points to allow the samples to exit early, thus reducing the cost of inference [for reducing computational resources of the at least one processor for generating the inference output,]. At each exit point, BranchyNet uses the entropy of a classification result (e.g., by softmax) as a measure of confidence in the prediction. If the entropy of a test sample is below a learned threshold value, meaning that the classifier is confident in the prediction [, and using, by the at least one processor, the inference output as a task result for a given task associated with the inference data.], the sample exits the network with the prediction result at this exit point, and is not processed by the higher network layers [generating an inference output by employing only the target sub-network of the NN on inference data].”).
While Teerapittayanon teaches a system that determines two subnetworks within a base neural network, Teerapittayanon does not explicitly teach:
wherein determining the first continuous sequence comprises randomly sampling the first continuous sequence from a predefined distribution over a sample space of continuous sequences of intermediate layers of the NN,
wherein determining the second continuous sequence comprises randomly sampling the second continuous sequence from the predefined distribution over the sample space of continuous sequences of intermediate layers of the NN,
Huang teaches:
wherein determining the first continuous sequence comprises randomly sampling the first continuous sequence from a predefined distribution over a sample space of continuous sequences of intermediate layers of the NN, (Huang, pg. 4, “To reduce the effective length of a neural network during training, we randomly skip layers entirely. We achieve this by introducing skip connections in the same fashion as ResNets, however the connection pattern is randomly altered for each mini batch. For each mini-batch we randomly select sets of layers [wherein determining the first continuous sequence comprises randomly sampling the first continuous sequence] and remove their corresponding transformation functions, only keeping the identity skip connection.”, and Huang, pg. 5, “We can achieve this goal by randomly dropping entire ResBlocks during training and bypassing their transformations through skip connections. Let b ∈ {0,1} denote a Bernoulli random variable, which indicates whether the lth ResBlock is active (b = 1) or inactive (b = 0). Further, let us denote the “survival” probability of ResBlock as p = Pr(b = 1) [from a predefined distribution over a sample space of continuous sequences of intermediate layers of the NN,].”).
wherein determining the second continuous sequence comprises randomly sampling the second continuous sequence from the predefined distribution over the sample space of continuous sequences of intermediate layers of the NN, (Huang, pg. 4, “To reduce the effective length of a neural network during training, we randomly skip layers entirely. We achieve this by introducing skip connections in the same fashion as ResNets, however the connection pattern is randomly altered for each mini batch. For each mini-batch we randomly select sets of layers [wherein determining the second continuous sequence comprises randomly sampling the second continuous sequence] and remove their corresponding transformation functions, only keeping the identity skip connection.”, and Huang, pg. 5, “We can achieve this goal by randomly dropping entire ResBlocks during training and bypassing their transformations through skip connections. Let b ∈ {0,1} denote a Bernoulli random variable, which indicates whether the lth ResBlock is active (b = 1) or inactive (b = 0). Further, let us denote the “survival” probability of ResBlock as p = Pr(b = 1) [from the predefined distribution over the sample space of continuous sequences of intermediate layers of the NN,].”).
Teerapittayanon and Huang are both in the same field of endeavor (i.e. machine learning). It would have been obvious for a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Teerapittayanon and Huang to teach the above limitation(s). The motivation for doing so is that stochastic depth training improves a deep neural network’s performance (cf. Huang, pg. 14, “Training with stochastic depth allows one to increase the depth of a network well beyond 1000 layers, and still obtain a reduction in test error. Because of its simplicity and practicality we hope that training with stochastic depth may become a new tool in the deep learning “toolbox”, and will help researchers scale their models to previously unattainable depths and capabilities.”).
Regarding claim 2, Teerapittayanon in view of Huang teaches the method of claim 1.
Huang further teaches:
wherein the method further comprises: during the first training iteration: determining a first depth index indicative of a first depth of the first continuous sequence of intermediate layers in the NN, and wherein the determining the first continuous sequence of intermediate layers includes: determining the continuous sequence of intermediate layers that is most adjacent to the input layer of the NN and which includes a total number of layers equal to the first depth index; (Huang, pg. 4 and see Figure 2, “Learning with stochastic depth is based on a simple intuition. To reduce the effective length of a neural network during training, we randomly skip layers entirely. We achieve this by introducing skip connections in the same fashion as ResNets, however the connection pattern is randomly altered for each minibatch; each minibatch is interpreted as different training iterations (i.e. wherein the method further comprises: during the first training iteration:). For each mini-batch we randomly select sets of layers [determining a first depth index indicative of a first depth of the first continuous sequence of intermediate layers in the NN,] and remove their corresponding transformation functions, only keeping the identity skip connection; Figure 2 shows that the random selection is done on layers after the input layer thus any set of randomly selected layers is interpreted as being most adjacent to the input layer (i.e. and wherein the determining the first continuous sequence of intermediate layers includes: determining the continuous sequence of intermediate layers that is most adjacent to the input layer of the NN).” and Huang, pg. 5, “We can achieve this goal by randomly dropping entire ResBlocks during training and bypassing their transformations through skip connections. Let b ∈ {0,1} denote a Bernoulli random variable, which indicates whether the lth ResBlock is active (b = 1) or inactive (b = 0). Further, let us denote the “survival” probability of ResBlock as p = Pr(b = 1) [and which includes a total number of layers equal to the first depth index;].”).
and wherein: during the second training iteration: determining a second depth index indicative of a second depth of the second continuous sequence of intermediate layers in the NN, the second depth index being different from the first depth index, and wherein the determining the second continuous sequence of intermediate layers includes: determining an other continuous sequence of intermediate layers that is most adjacent to the input layer of the NN and which includes a total number of layers equal to the second depth index. (Huang, pg. 4 and see Figure 2, “Learning with stochastic depth is based on a simple intuition. To reduce the effective length of a neural network during training, we randomly skip layers entirely. We achieve this by introducing skip connections in the same fashion as ResNets, however the connection pattern is randomly altered for each minibatch; each minibatch is interpreted as different training iterations (i.e. and wherein: during the second training iteration:). For each mini-batch we randomly select sets of layers [determining a second depth index indicative of a second depth of the second continuous sequence of intermediate layers in the NN,] and remove their corresponding transformation functions, only keeping the identity skip connection; Figure 2 shows that the random selection is done on layers after the input layer thus any set of randomly selected layers is interpreted as being most adjacent to the input layer (i.e. and wherein the determining the second continuous sequence of intermediate layers includes: determining an other continuous sequence of intermediate layers that is most adjacent to the input layer of the NN).” and Huang, pg. 5, “We can achieve this goal by randomly dropping entire ResBlocks during training and bypassing their transformations through skip connections. Let b ∈ {0,1} denote a Bernoulli random variable, which indicates whether the lth ResBlock is active (b = 1) or inactive (b = 0). Further, let us denote the “survival” probability of ResBlock as p = Pr(b = 1) [and which includes a total number of layers equal to the second depth index.].” and Huang, pg. 4, “Similar to Dropout, stochastic depth can be interpreted as training an ensemble of networks, but with different depths, possibly achieving higher diversity among ensemble members than ensembling those with the same depth [the second depth index being different from the first depth index,].”).
It would have been obvious to one of ordinary skill in the art before the effective filling date of the present application to combine the teachings of Huang with the teachings of Teerapittayanon for the same reasons disclosed in claim 1.
Regarding claim 3, Teerapittayanon in view of Huang teaches the method of claim 2.
Huang further teaches:
wherein randomly sampling the first continuous sequence comprises randomly determining the first depth index from an interval of depth indexes, and randomly sampling the second continuous sequence comprises randomly determining the second depth index from the interval of depth indexes, (Huang, pg. 4, “To reduce the effective length of a neural network during training, we randomly skip layers entirely. We achieve this by introducing skip connections in the same fashion as ResNets, however the connection pattern is randomly altered for each mini batch. For each mini-batch we randomly select sets of layers [wherein randomly sampling the first continuous sequence comprises randomly determining the first depth index…and randomly sampling the second continuous sequence comprises randomly determining the second depth index] and remove their corresponding transformation functions, only keeping the identity skip connection.”, and Huang, pg. 5, “We can achieve this goal by randomly dropping entire ResBlocks during training and bypassing their transformations through skip connections. Let b ∈ {0,1} denote a Bernoulli random variable, which indicates whether the lth ResBlock is active (b = 1) or inactive (b = 0). Further, let us denote the “survival” probability of ResBlock as p = Pr(b = 1) […from an interval of depth indexes,].”).
the interval of depth indexes having been pre-determined based on a depth of the NN. (Huang, pg. 4, “Similar to Dropout, stochastic depth can be interpreted as training an ensemble of networks, but with different depths, possibly achieving higher diversity among ensemble members than ensembling those with the same depth. Different from Dropout, we make the network shorter [the interval of depth indexes having been pre-determined based on a depth of the NN.]”).
It would have been obvious to one of ordinary skill in the art before the effective filling date of the present application to combine the teachings of Huang with the teachings of Teerapittayanon for the same reasons disclosed in claim 1.
Regarding claim 6, Teerapittayanon in view of Huang teaches the method of claim 1. Teerapittayanon further teaches wherein the selecting the target sub-network comprises comparing at least one of:
(i) a first accuracy parameter of the first sub-network and a second accuracy parameter of the second sub-network,
(ii) a first latency parameter of the first sub-network and a second latency parameter of the second sub-network, and
(iii) a first importance parameter of the first sub-network and a second importance parameter of the second sub-network.
(Teerapittayanon, pg. 2465 col. 1, “At each exit point, BranchyNet uses the entropy of a classification result (e.g., by softmax) as a measure of confidence in the prediction. If the entropy of a test sample is below a learned threshold value, meaning that the classifier is confident in the prediction, the sample exits the network with the prediction result at this exit point, and is not processed by the higher network layers. If the entropy value is above the threshold, then the classifier at this exit point is deemed not confident, and the sample continues to the next exit point in the network [wherein the selecting the target sub-network comprises comparing at least one of: (i) a first accuracy parameter of the first sub-network and a second accuracy parameter of the second sub-network,].”).
Regarding claim 7, Teerapittayanon in view of Huang teaches the method of claim 1. Huang further teaches wherein the plurality of intermediate layers is a plurality of architectural blocks of the NN, a given one of the plurality of architectural blocks including a sub-set of intermediate layers for generating an output of the given one of the plurality of architectural blocks. (Huang, pg. 5, “Stochastic depth aims to shrink the depth of a network during training, while keeping it unchanged during testing. We can achieve this goal by randomly dropping entire ResBlocks during training and bypassing their transformations through skip connections [wherein the plurality of intermediate layers is a plurality of architectural blocks of the NN, a given one of the plurality of architectural blocks including a sub-set of intermediate layers for generating an output of the given one of the plurality of architectural blocks.].”).
It would have been obvious to one of ordinary skill in the art before the effective filling date of the present application to combine the teachings of Huang with the teachings of Teerapittayanon for the same reasons disclosed in claim 1.
Regarding claim 8, Teerapittayanon in view of Huang teaches the method of claim 7. Huang further teaches wherein the plurality of architectural blocks include at least one of:
(i) a convolutional block with at least one convolutional layer;
(ii) a pooling block with at least one pooling layer;
(iii) a fully connected block with at least one fully-connected layer;
(iv) a residual block with at least one skip connection;
(v) a batch normalization block with at least one batch normalization layer;
(vi) a recurrent block with at least one recurrence mechanism;
(vii) an attention block with at least one self-attention mechanism; and
(viii) an activation block with at least one activation layer.
(Huang, pg. 5, “Stochastic depth aims to shrink the depth of a network during training, while keeping it unchanged during testing. We can achieve this goal by randomly dropping entire ResBlocks during training and bypassing their transformations through skip connections [wherein the plurality of architectural blocks include at least one of:…(iv) a residual block with at least one skip connection;].”).
It would have been obvious to one of ordinary skill in the art before the effective filling date of the present application to combine the teachings of Huang with the teachings of Teerapittayanon for the same reasons disclosed in claim 1.
Regarding claim 9, Teerapittayanon in view of Huang teaches the method of claim 1. Teerapittayanon further teaches wherein the output layer is at least two output layers, and wherein the output layer of the first sub-network is a first one from the at least two output layers, and the output layer of the second sub-network is a second one from the at least two output layers, the first and second one of the at least two output layers being different output layers. (Teerapittayanon, pg. 2466 col. 1 and see Figure 1, “A branch is a subset of the network containing contiguous layers, which do not overlap other branches; since the branches do not overlap, each exit is interpreted as having their own different output layers (i.e. the first and second one of the at least two output layers being different output layers.), followed by an exit point. The main branch can be considered the baseline (original) network before side branches are added; Figure 1 shows that there are 3 exits of the model which is interpreted as having 3 output layers as they perform classification at each exit (i.e. wherein the output layer is at least two output layers, and wherein the output layer of the first sub-network is a first one from the at least two output layers, and the output layer of the second sub-network is a second one from the at least two output layers,).”).
Regarding claim 10, the claim is similar to claim 1 and rejected under the same rationales. Teerapittayanon further teaches the additional limitations …the system comprising a at least a processor and a memory storing a plurality of executable instructions which, when executed by the at least a processor, cause the system to:… (Teerapittayanon, pg. 2467 col. 1, “We evaluate Branchy-LeNet (B-LeNet) on the MNIST dataset and both Branchy-AlexNet (B-AlexNet) and Branchy-ResNet (B-ResNet) on the CIFAR10 dataset. We present evaluation results for both CPU and GPU. We use a 3.0 GHz CPU with 20MB L3 Cache and NVIDIA GeForce GTX TITANX (Maxwell) 12GB GPU […the system comprising a at least a processor and a memory storing a plurality of executable instructions which, when executed by the at least a processor, cause the system to:…].”).
Regarding claims 11-12, the claims are similar to claims 2-3 and are rejected under the same rationales.
Regarding claims 15-18, these claims are similar to claims 6-9 and are rejected under the same rationales.
Claims 4 and 13 are rejected under 35 U.S.C. 103 as being unpatentable over Teerapittayanon, et al., “BranchyNet: Fast Inference via Early Exiting from Deep Neural Networks” (“Teerapittayanon”) in view of Huang, et al., Non-Patent Literature “Deep Networks with Stochastic Depth” (“Huang”) and further in view of Yu, et al., Non-Patent Literature “Slimmable Neural Networks” (“Yu”).
Regarding claim 4, Teerapittayanon in view of Huang teaches the method of claim 1. While the combination teaches a system that determines two subnetworks within a base neural network at different depths, the combination does not explicitly teach wherein the method further comprises: during the first training iteration: determining a first width index for the first continuous sequence of intermediate layers indicative of a first partial width of the first continuous sequence of intermediate layers to be trained during the first training iteration, and wherein the training the first sub-network includes: training only the first partial width of the first continuous sequence of intermediate layers based on the training data, and wherein: during the second training iteration: determining a second width index for the second continuous sequence of intermediate layers indicative of a second partial width of the second continuous sequence of intermediate layers to be trained during the second training iteration, the second width index being different from the first width index, the training the second sub-network includes: training only the second partial width of the second continuous sequence of intermediate layers based on the training data.
Yu teaches:
wherein the method further comprises: during the first training iteration: determining a first width index for the first continuous sequence of intermediate layers indicative of a first partial width of the first continuous sequence of intermediate layers to be trained during the first training iteration, (Yu, pg. 2, “The parameters of all model variants are shared and the active channels in different layers can be adjusted. For brevity, we denote a model variant in a slimmable network as a switch [wherein the method further comprises: during the first training iteration: determining a first width index for the first continuous sequence of intermediate layers indicative of a first partial width of the first continuous sequence of intermediate layers to be trained], the number of active channels in a switch as its width. 0.25× represents that the width in all layers are scaled by 0.25 of the full model.”, and Yu, pg. 4, “Motivated by the investigations above, we present a simple and highly effective approach, named Switchable Batch Normalization (S-BN), that employs independent batch normalization (Ioffe & Szegedy, 2015) for different switches in a slimmable network [to be trained during the first training iteration,].”, and Yu, pg. 4, “The switchable width list [width index] is predefined, indicating the available switches in a slimmable network.”).
and wherein the training the first sub-network includes: training only the first partial width of the first continuous sequence of intermediate layers based on the training data, (Yu, pg. 4, “Motivated by the investigations above, we present a simple and highly effective approach, named Switchable Batch Normalization (S-BN), that employs independent batch normalization (Ioffe & Szegedy, 2015) for different switches in a slimmable network [and wherein the training the first sub-network includes: training only the first partial width of the first continuous sequence of intermediate layers based on the training data,].”).
and wherein: during the second training iteration: determining a second width index for the second continuous sequence of intermediate layers indicative of a second partial width of the second continuous sequence of intermediate layers to be trained during the second training iteration, the second width index being different from the first width index, (Yu, pg. 2, “The parameters of all model variants are shared and the active channels in different layers can be adjusted. For brevity, we denote a model variant in a slimmable network as a switch [and wherein: during the second training iteration: determining a second width index for the second continuous sequence of intermediate layers indicative of a second partial width of the second continuous sequence of intermediate layers], the number of active channels in a switch as its width. 0.25× represents that the width in all layers are scaled by 0.25 of the full model.”, and Yu, pg. 4, “Motivated by the investigations above, we present a simple and highly effective approach, named Switchable Batch Normalization (S-BN), that employs independent batch normalization (Ioffe & Szegedy, 2015) for different switches [the second width index being different from the first width index,] in a slimmable network [to be trained during the second training iteration,].”, and Yu, pg. 4, “The switchable width list [width index] is predefined, indicating the available switches in a slimmable network.”).
the training the second sub-network includes: training only the second partial width of the second continuous sequence of intermediate layers based on the training data. (Yu, pg. 4, “Motivated by the investigations above, we present a simple and highly effective approach, named Switchable Batch Normalization (S-BN), that employs independent batch normalization (Ioffe & Szegedy, 2015) for different switches in a slimmable network [the training the second sub-network includes: training only the second partial width of the second continuous sequence of intermediate layers based on the training data.].”).
Teerapittayanon, in view of Huang, and Yu are both in the same field of endeavor (i.e. machine learning). It would have been obvious for a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Teerapittayanon, in view of Huang, and Yu to teach the above limitation(s). The motivation for doing so is that having adaptable widths allows models to be adapted for different resource budgets (cf. Yu, pg. 2, “The question remains: Given budgets of resources, how to instantly, adaptively and efficiently trade off between accuracy and latency for neural networks at runtime? In this work we introduce slimmable neural networks, a new class of networks executable at different widths, as a general solution to trade off between accuracy and latency on the fly.”).
Regarding claim 13, the claim is similar to claim 4 and is rejected under the same rationales.
Claims 5 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Teerapittayanon, et al., “BranchyNet: Fast Inference via Early Exiting from Deep Neural Networks” (“Teerapittayanon”) in view of Huang, et al., Non-Patent Literature “Deep Networks with Stochastic Depth” (“Huang”) and further in view of Yu, et al., Non-Patent Literature “Slimmable Neural Networks” (“Yu”) and Han, et al., Foreign Patent Publication CN112418392A (“Han”), please use provided translated copy for claim mapping.
Regarding claim 5, Teerapittayanon in view of Huang and Yu teaches the method of claim 4. Yu further teaches the interval of width indexes having been pre-determined based on a width of the plurality of intermediate layers of the NN. (Yu, pg. 4, “The switchable width list is predefined, indicating the available switches in a slimmable network [the interval of width indexes having been pre-determined based on a width of the plurality of intermediate layers of the NN.].”).
It would have been obvious to one of ordinary skill in the art before the effective filling date of the present application to combine the teachings of Yu with the teachings of Teerapittayanon and Huang for the same reasons disclosed in claim 4.
While the combination teaches a system that determines two subnetworks within a base neural network with a width index, the combination does not explicitly teach wherein the determining the first width index includes randomly determining the first width index from an interval of width indexes, and the determining the second width index includes randomly determining the second width index from the interval of width indexes,
Han teaches wherein the determining the first width index includes randomly determining the first width index from an interval of width indexes, and the determining the second width index includes randomly determining the second width index from the interval of width indexes, (Han, pg. 15, “The sampling mode in the search space may be random sampling, or sampling according to distribution, which may be specifically adjusted according to the actual application scenario, and this application does not limit this…For example, in the case of sampling the width, the sampling probability may be determined from the distribution of the width, and the sampling probability may be larger for a wide range in which a large number of distributions are distributed, and may be smaller for a wide range in which a small number of distributions are distributed [wherein the determining the first width index includes randomly determining the first width index from an interval of width indexes, and the determining the second width index includes randomly determining the second width index from the interval of width indexes,].”).
Teerapittayanon, in view of Huang and Yu, and Han are both in the same field of endeavor (i.e. machine learning). It would have been obvious for a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Teerapittayanon, in view of Huang and Yu, and Han to teach the above limitation(s). The motivation for doing so is that using random sampling removes bias in searching for model parameters (cf. Han, see pg. 15).
Regarding claim 14, the claim is similar to claim 5 and is rejected under the same rationales.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Kim, et al., “Nestednet: Learning nested sparse structures in deep neural networks” teaches a nested sparse network that consists of multiple levels of networks with a different sparsity at each level, and higher level networks share parameters with lower level networks to enable nested learning.
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 NICHOLAS S WU whose telephone number is (571)270-0939. The examiner can normally be reached Monday - Friday 8:00 am - 4:00 pm EST.
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, Michelle Bechtold can be reached at 571-431-0762. 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.
/N.S.W./Examiner, Art Unit 2148 /MICHELLE T BECHTOLD/Supervisory Patent Examiner, Art Unit 2148
Prosecution Insights