DETAILED ACTION
1. This communication is in response to the amendments filed on January 8, 2026 for Application No. 17/932,104 in which Claims 1, 3-8, 10-11, 13-18, and 21-22 are presented for examination.
Notice of Pre-AIA or AIA Status
2. 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
3. The amendments filed on January 8, 2026 have been considered. Claims 1, 3-8, 10-11, and 16 have been amended. Claims 2, 9, 12, and 19-20 have been cancelled. Claims 21-22 are newly added. Thus, Claims 1, 3-8, 10-11, 13-18, and 21-22 are pending and presented for examination.
4. Applicant’s arguments filed January 8, 2026 with respect to the 35 U.S.C. 112(b) rejection have been fully considered and are persuasive. Thus, the 35 U.S.C. 112(b) rejection has been withdrawn.
5. Applicant’s arguments filed January 8, 2026 with respect to the 35 U.S.C. 101 software per se rejection have been fully considered and are persuasive. Thus, the 35 U.S.C. 101 rejection has been withdrawn.
6. Applicant’s arguments filed January 8, 2026 with respect to the 35 U.S.C. 102/103 rejections have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument. Note: Examiner has replaced the secondary reference of record, Rastegari et al., with the Helwegen reference (US PG-PUB 20220405576) to teach the amended & argued limitations – thus, Applicant’s arguments are moot.
Claim Rejections - 35 USC § 103
7. 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.
8. Claims 1, 3-4, 10-11, 13-14, 16, and 21-22 are rejected under 35 U.S.C. 103 as being unpatentable over Courbariaux et al. (hereinafter Courbariaux) (“Binarized Neural Networks: Training Neural Networks with Weights and Activations Constrained to +1 or -1”), in view of Helwegen (hereinafter Helwegen) (US PG-PUB 20220405576).
Regarding Claim 1, Courbariaux teaches a computing system comprising a processor and a non-transitory computer-readable medium storing executable computer program instructions that cause the processor to operate (While Courbariaux Pg. 6 Section 3. Very Power Efficient in Forward Pass, broadly recites the use of computer hardware comprising both general-purpose or specialized composed of memories, arithmetic operators, and control logics, Courbariaux does not explicitly disclose a computing system comprising a processor and non-transitory computer-readable medium. See introduction of Helwegen reference below for teaching of a computing system comprising a processor and non-transitory computer-readable medium) a robust and accurate binary neural network (RA-BNN) (Courbariaux, Pg. 1, Abstract, “We introduce a method to train Binarized Neural Networks (BNNs)- neural networks with binary weights and activations at run-time. At training-time the binary weights and activations are used for computing the parameters gradients. During the forward pass, BNNs drastically reduce memory size and accesses, and replace most arithmetic operations with bit-wise operations, which is expected to substantially improve power-efficiency.”, therefore, a robust and accurate binary neural network is disclosed), wherein the RA-BNN comprises:
a first deep neural network (DNN) layer having a non-binary input and binarized weights (Courbariaux, Pg. 4, “In a BNN, only the binarized values of the weights and activations are used in all calculations. As the output of one layer is the input of the next, all the layers inputs are binary, with the exception of the first layer.”, thus, weights and inputs of each layer of the binary neural network are binarized. Further, the input of the first layer is not binarized);
a last DNN layer is configured to accept a non-binary input and binarized weights (See introduction of Helwegen reference below for teaching of a last DNN layer configured to accept a non-binary input and binarized weights); and
one or more intermediate DNN layers between the first DNN layer and the last DNN layer, wherein the one or more intermediate DNN layers have binary inputs and binarized weights (Courbariaux, Pg. 4, “In a BNN, only the binarized values of the weights and activations are used in all calculations. As the output of one layer is the input of the next, all the layers inputs are binary, with the exception of the first layer.”, thus, weights and inputs of each layer of the binary neural network are binarized. This is also supported by the abstract of Courbariaux which states that the weights of the binary neural network are all binarized).
While Courbariaux Pg. 6 Section 3. Very Power Efficient in Forward Pass, broadly recites the use of computer hardware comprising both general-purpose or specialized composed of memories, arithmetic operators, and control logics, Courbariaux does not explicitly disclose a computing system comprising a processor and non-transitory computer-readable medium
However, Helwegen teaches a computing system (Helwegen, Figure 5, label 500 disclosing a computing device) comprising a processor and non-transitory computer-readable medium storing executable computer program instructions that cause the processor (Helwegen, Par. [0188], “The memory 530 is configured to store instructions and data for utilization by the processor 520. The memory 530 may be a non-transitory volatile memory device, such as a random access memory (RAM) device.”, therefore, a computing device (Figure 5, label 500) comprising a processor (label 520) and a non-transitory computer-readable medium storing instructions (label 530 or label 540)) to […]
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the robust and accurate binary neural network (RA-BNN), as disclosed by Courbariaux to include a computing system comprising a processor and non-transitory computer-readable medium storing executable computer program instructions that cause the processor to operate the RA-BNN, as disclosed by Helwegen. One of ordinary skill in the art would have been motivated to make this modification to enable the use of a computing device comprising processor and memory, which may efficiently implement and operate the neural network (Helwegen, Par. [0133], “FIG. 3A shows a method 300A for classifying an image using a neural network. The neural network may be the neural network 100 described in relation to FIG. 1 . The method 300A may be implemented as one or more computer-executable instructions executed on one or more computing devices, e.g. the computing device 500 described in relation to FIG. 5.”)
Courbariaux does not explicitly disclose a last DNN layer is configured to accept a non-binary input and binarized weights.
However, Helwegen teaches a last DNN layer is configured to accept a non-binary input and binarized weights (Helwegen, Par. [0106], “The combined intermediate output is received by a set of higher-precision layers 160, which are layers comprising higher precision values for the inputs (activations) and the outputs of the layer. The set of higher-precision layers 160 calculate a neural network output 170 based on the combined intermediate output. The higher precision neural network layers 160 may comprise higher precision weights or ultra-low precision weights for example.”, thus, the last DNN layer (which produces the neural network output as depicted by Figure 1) is configured to be a higher-precision layer, meaning that the layer accepts a non-binary input (higher precision values for the inputs/activations) and binarized weights (ultra-low precision weights, which may comprise binarized weights as supported by Helwegen Par. [0166]))
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the computing system to operate a robust and accurate binary neural network, as disclosed by Courbariaux in view of Helwegen to include a last DNN layer is configured to accept a non-binary input and binarized weights, as disclosed by Helwegen. One of ordinary skill in the art would have been motivated to make this modification to preserve numerical precision whilst also maintaining efficiency on resource-constrained devices (Helwegen, Par. [0035], “Such ultra-low precision neural network layers may use smaller storage capacity compared to higher-precision neural network layers. In particular, storing ultra-low precision weights may require a smaller amount of non-volatile memory, while ultra-low precision activations may require a smaller amount of working memory. Furthermore, performing computations using ultra-low precision weights and activations may be more energy-efficient than using high-precision values on commodity hardware and on specialized hardware, while the increase in energy efficiency on specialized hardware may be particularly pronounced.” & Par. [0059], “The first intermediate output is a higher-precision output. The reduced first intermediate output is a higher-precision output. Thus, numerical, or precision, information is preserved in the reduced first intermediate output.”)
Regarding Claim 3, Courbariaux in view of Helwegen teaches the computing system of claim 1, wherein the processor is configured to train the RA-BNN using early growth (Courbariaux, Pg. 1, Abstract, “We introduce a method to train Binarized Neural Networks (BNNs)- neural networks with binary weights and activations at run-time. At training-time the binary weights and activations are used for computing the parameters gradients. During the forward pass, BNNs drastically reduce memory size and accesses, and replace most arithmetic operations with bit-wise operations, which is expected to substantially improve power-efficiency.”, therefore, a method for training a BNN using early growth is disclosed, as per Applicant’s definition of “early growth” supported by instant dependent claim 4).
Regarding Claim 4, Courbariaux in view of Helwegen teaches the computing system of claim 3, wherein processor is configured to train the RA-BNN using early growth by:
training and channel-wise growing the RA-BNN from an initial RA-BNN to a larger RA-BNN (Courbariaux, Pg. 7, “When using a ConvNet architecture with binary weights, the number of unique filters is bounded by the filter size. For example, in our implementation we use filters of size 3 3, so the maximum number of unique 2D filters is 29 = 512. However, this should not prevent expanding the number of feature maps beyond this number, since the actual filter is a 3D matrix.”, therefore, the binary neural network is trained and the number of feature maps/channels may grow/be increased as training continues); and
retraining the larger RA-BNN to minimize a defined loss (Courbariaux, Pg. 5, Figure 1, which depicts the training/retraining of the BNN across different data sets and minimizes a defined loss/square hinge loss).
Regarding Claim 10, Courbariaux in view of Helwegen teaches the computing system of claim 1, wherein the RA-BNN further comprises a channel index for each layer (Helwegen, Par. [0085], “An example of a binarized convolutional neural network layer for an input having 2 spatial dimensions and a number of channels is described mathematically below. A binarized convolutional neural network layer comprises a number of filters. The number of output channels from the layer corresponds to the number of filters in the layer. Each filter has the same depth as the input to the layer, i.e. the number of input channels. Each filter is moved along the width and height of the input.”, therefore, a channel index may be associated for each layer of the neural network)
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the computing system of claim 1, as disclosed by Courbariaux in view of Helwegen to include wherein the RA-BNN further comprises a channel index for each layer, as disclosed by Helwegen. One of ordinary skill in the art would have been motivated to make this modification to enable efficient and accurate feature representation across a plurality of different features as related to image processing, by utilizing channels/a channel index for each layer (Helwegen, Par. [0046], “A (2D) matrix input 110 may be described as being composed of multiple channels, with each channel being a vector. Hence, the dimensions of a (2D) matrix input may be described as one spatial dimension (the length of each vector), and a number of channels. In the case of such an input representing audio data, the spatial dimension may be the number of audio timesteps represented in the input, with the number of channels being the number of different features, e.g. the number of frequency bands. In the case of a (2D) matrix input representing sensor data, the spatial dimension may be the number of timesteps represented in the input, with the number of channels being the number of different readings being captured, the number of different aspects being captured, or a combination thereof.”).
Regarding Claim 11, Courbariaux teaches a method for strengthening a binary neural network (BNN) against adversarial noise injection (Courbariaux, Pg. 1, Abstract, “We introduce a method to train Binarized Neural Networks (BNNs)- neural networks with binary weights and activations at run-time. At training-time the binary weights and activations are used for computing the parameters gradients. During the forward pass, BNNs drastically reduce memory size and accesses, and replace most arithmetic operations with bit-wise operations, which is expected to substantially improve power-efficiency.”, therefore, methods for strengthening a binary neural network are disclosed. Further, Courbariaux Pg. 2 Section 1. Binarized Neural Networks further describes how noise may be “averaged out” utilizing stochastic gradient contributions), the method comprising:
binarizing weights of each layer of the BNN; and binarizing inputs of each intermediate layer of the BNN between a first layer and a last layer such that an input of the first layer and the last layer (See introduction of Helwegen reference below for explicit teaching of wherein an input of the last layer is not binarized) are not binarized (Courbariaux, Pg. 4, “In a BNN, only the binarized values of the weights and activations are used in all calculations. As the output of one layer is the input of the next, all the layers inputs are binary, with the exception of the first layer.”, thus, weights and inputs of each layer of the binary neural network are binarized. Further, the input of the first layer is not binarized).
Courbariaux does not explicitly disclose binarizing inputs […] such that an input of the […] last layer is not binarized
However, Helwegen teaches binarizing inputs […] such that an input of the […] last layer is not binarized (Helwegen, Par. [0106], “The combined intermediate output is received by a set of higher-precision layers 160, which are layers comprising higher precision values for the inputs (activations) and the outputs of the layer. The set of higher-precision layers 160 calculate a neural network output 170 based on the combined intermediate output. The higher precision neural network layers 160 may comprise higher precision weights or ultra-low precision weights for example.”, thus, the last DNN layer (which produces the neural network output as depicted by Figure 1) is configured to be a higher-precision layer, meaning that the layer accepts a non-binary input (higher precision values for the inputs/activations) and binarized weights (ultra-low precision weights, which may comprise binarized weights as supported by Helwegen Par. [0166]))
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the method for strengthening a binary neural network (BNN) against adversarial noise injection, as disclosed by Courbariaux to include binarizing inputs […] such that an input of the […] last layer is not binarized, as disclosed by Helwegen. One of ordinary skill in the art would have been motivated to make this modification to preserve numerical precision whilst also maintaining efficiency on resource-constrained devices (Helwegen, Par. [0035], “Such ultra-low precision neural network layers may use smaller storage capacity compared to higher-precision neural network layers. In particular, storing ultra-low precision weights may require a smaller amount of non-volatile memory, while ultra-low precision activations may require a smaller amount of working memory. Furthermore, performing computations using ultra-low precision weights and activations may be more energy-efficient than using high-precision values on commodity hardware and on specialized hardware, while the increase in energy efficiency on specialized hardware may be particularly pronounced.” & Par. [0059], “The first intermediate output is a higher-precision output. The reduced first intermediate output is a higher-precision output. Thus, numerical, or precision, information is preserved in the reduced first intermediate output.”)
Regarding Claim 13, Courbariaux teaches the method of claim 11, further comprising training and channel-wise growing the BNN from an initial BNN to a larger BNN (Courbariaux, Pg. 7, “When using a ConvNet architecture with binary weights, the number of unique filters is bounded by the filter size. For example, in our implementation we use filters of size 3 3, so the maximum number of unique 2D filters is 29 = 512. However, this should not prevent expanding the number of feature maps beyond this number, since the actual filter is a 3D matrix.”, therefore, the binary neural network is trained and the number of feature maps/channels may grow/be increased as training continues).
Regarding Claim 14, Courbariaux teaches the method of claim 13, further comprising retraining the larger BNN to minimize a defined loss (Courbariaux, Pg. 5, Figure 1, which depicts the training/retraining of the BNN across different data sets and minimizes a defined loss/square hinge loss).
Regarding Claim 16, Courbariaux teaches a method for training a binary neural network (BNN) using early growth (Courbariaux, Pg. 1, Abstract, “We introduce a method to train Binarized Neural Networks (BNNs)- neural networks with binary weights and activations at run-time. At training-time the binary weights and activations are used for computing the parameters gradients. During the forward pass, BNNs drastically reduce memory size and accesses, and replace most arithmetic operations with bit-wise operations, which is expected to substantially improve power-efficiency.”, therefore, a method for training a BNN using early growth is disclosed), the method comprising:
binarizing weights of each layer of the BNN; binarizing inputs of each intermediate layer of the BNN between a first layer and a last layer such that an input of the first layer and the last layer (See introduction of Helwegen reference below for explicit teaching of wherein an input of the last layer is not binarized) are not binarized (Courbariaux, Pg. 4, “In a BNN, only the binarized values of the weights and activations are used in all calculations. As the output of one layer is the input of the next, all the layers inputs are binary, with the exception of the first layer.”, thus, weights and inputs of each layer of the binary neural network are binarized. Further, the input of the first layer is not binarized);
training and channel-wise growing the BNN from an initial BNN to a larger BNN (Courbariaux, Pg. 7, “When using a ConvNet architecture with binary weights, the number of unique filters is bounded by the filter size. For example, in our implementation we use filters of size 3 3, so the maximum number of unique 2D filters is 29 = 512. However, this should not prevent expanding the number of feature maps beyond this number, since the actual filter is a 3D matrix.”, therefore, the binary neural network is trained and the number of feature maps/channels may grow/be increased as training continues); and
retraining the larger BNN to minimize a defined loss (Courbariaux, Pg. 5, Figure 1, which depicts the training/retraining of the BNN across different data sets and minimizes a defined loss/square hinge loss).
Courbariaux does not explicitly disclose binarizing inputs […] such that an input of the […] last layer is not binarized
However, Helwegen teaches binarizing inputs […] such that an input of the […] last layer is not binarized (Helwegen, Par. [0106], “The combined intermediate output is received by a set of higher-precision layers 160, which are layers comprising higher precision values for the inputs (activations) and the outputs of the layer. The set of higher-precision layers 160 calculate a neural network output 170 based on the combined intermediate output. The higher precision neural network layers 160 may comprise higher precision weights or ultra-low precision weights for example.”, thus, the last DNN layer (which produces the neural network output as depicted by Figure 1) is configured to be a higher-precision layer, meaning that the layer accepts a non-binary input (higher precision values for the inputs/activations) and binarized weights (ultra-low precision weights, which may comprise binarized weights as supported by Helwegen Par. [0166]))
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the method for training a binary neural network (BNN) using early growth, as disclosed by Courbariaux to include binarizing inputs […] such that an input of the […] last layer is not binarized, as disclosed by Helwegen. One of ordinary skill in the art would have been motivated to make this modification to preserve numerical precision whilst also maintaining efficiency on resource-constrained devices (Helwegen, Par. [0035], “Such ultra-low precision neural network layers may use smaller storage capacity compared to higher-precision neural network layers. In particular, storing ultra-low precision weights may require a smaller amount of non-volatile memory, while ultra-low precision activations may require a smaller amount of working memory. Furthermore, performing computations using ultra-low precision weights and activations may be more energy-efficient than using high-precision values on commodity hardware and on specialized hardware, while the increase in energy efficiency on specialized hardware may be particularly pronounced.” & Par. [0059], “The first intermediate output is a higher-precision output. The reduced first intermediate output is a higher-precision output. Thus, numerical, or precision, information is preserved in the reduced first intermediate output.”)
Regarding Claim 21, Courbariaux in view of Helwegen teaches the computing system of claim 1, wherein the first DNN layer, the one or more intermediate DNN layers, and the last DNN comprise floating point outputs (Helwegen, Par. [0106], “The combined intermediate output is received by a set of higher-precision layers 160, which are layers comprising higher precision values for the inputs (activations) and the outputs of the layer. The set of higher-precision layers 160 calculate a neural network output 170 based on the combined intermediate output.”, thus, the layers of the DNN may comprise floating point outputs).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the computing system of claim 1, as disclosed by Courbariaux in view of Helwegen to include wherein the first DNN layer, the one or more intermediate DNN layers, and the last DNN comprise floating point outputs, as disclosed by Helwegen. One of ordinary skill in the art would have been motivated to make this modification to preserve numerical precision, hence improving numerical accuracy of computations (Helwegen, Par. [0059], “The first intermediate output is a higher-precision output. The reduced first intermediate output is a higher-precision output. Thus, numerical, or precision, information is preserved in the reduced first intermediate output.”)
Regarding Claim 22, Courbariaux in view of Helwegen teaches the computing system of claim 21, wherein the processor is configured to binarize the floating point outputs of the first DNN layer and the one or more intermediate DNN layers (Helwegen, Par. [0205], “The intermediate output of the higher-precision convolutional layer 622-2 is received by a quantization layer 642-1. The quantization layer 642-1 is a precision-reducing layer which binarizes the intermediate output of the higher-precision convolutional layer 622-2 to obtain 1-bit values each representing 1 or −1. The quantization layer 642-1 calculates a binarized intermediate output with the same dimensions as the intermediate output of the higher-precision convolutional layer 622-2, 48×64×32, with each value therein being a binary value representing either 1 or −1.”, thus, the processor may be configured to binarize floating point outputs of the DNN layers)
The reasons of obviousness have been noted in the rejection of Claim 21 above and applicable herein.
9. Claims 5-8 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Courbariaux et al. (hereinafter Courbariaux) (“Binarized Neural Networks: Training Neural Networks with Weights and Activations Constrained to +1 or -1”), in view of Helwegen (hereinafter Helwegen) (US PG-PUB 20220405576), further in view of Csordas et al. (hereinafter Csordas) (“Are Neural Nets Modular? Inspecting Functional Modularity through Differentiable Weight Masks”).
Regarding Claim 5, Courbariaux in view of Helwegen teaches the computing system of claim 3.
Courbariaux in view of Helwegen does not explicitly disclose wherein the early growth comprises: at least one learning binary mask associated with at least one weight channel.
However, Csordas teaches wherein the early growth comprises: at least one learning binary mask associated with at least one weight channel (Csordas, Pg. 3, “Training is done by applying the loss function defined by the target task and backpropagating into the logits li. Typically multiple (between 4–8) binary masks are sampled and applied to different parts of a batch to improve the quality of the estimated gradient.”, therefore, the training comprises early growth wherein at least one learning binary mask is associated with a weight channel).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the computing system of claim 3 wherein the processor is configured to train the RA-BNN using early growth, as disclosed by Courbariaux in view of Helwegen to include wherein the early growth comprises: at least one learning binary mask associated with at least one weight channel, as disclosed by Csordas. One of ordinary skill in the art would have been motivated to make this modification to enable the use of binary weight masks which may identify individual weights responsible for specific functions, hence improving system accuracy and performance for a target function (Csordas, Pg. 1, “In this paper, we present a novel method based on learning binary weight masks to identify individual weights and subnets responsible for specific functions. Using this powerful tool, we contribute an extensive study of emerging modularity in NNs that covers several standard architectures and datasets. We demonstrate how common NNs fail to reuse submodules and offer new insights into the related issue of systematic generalization on language tasks.”)
Regarding Claim 6, Courbariaux in view of Helwegen teaches the computing system of claim 3.
Courbariaux in view of Helwegen does not explicitly disclose wherein the early growth starts from a given baseline model and each channel is associated with a trainable mask.
However, Csordas teaches wherein the early growth starts from a given baseline model and each channel is associated with a trainable mask (Csordas, Pg. 1, “Given an auxiliary task corresponding to a particular target function of interest (e.g., train only on a specific subset of the samples from the original dataset), we train probabilistic, binary, but differentiable masks for all weights (while the NN’s weights remain frozen). The result is a binary mask exhibiting the module necessary to perform the target function”, therefore, each channel is associated with a trainable differentiable mask).
The reasons of obviousness have been noted in the rejection of Claim 5 above and applicable herein.
Regarding Claim 7, Courbariaux in view of Helwegen in view of Csordas teaches the computing system of claim 6, wherein the processor is configured to create an output filter channel when the mask switches from 0 to 1 for the first time (Csordas, Pg. 3, “Training is done by applying the loss function defined by the target task and backpropagating into the logits li. Typically multiple (between 4–8) binary masks are sampled and applied to different parts of a batch to improve the quality of the estimated gradient. […] At the end of the training process, deterministic binary masks Mi ∈ {0,1} for weights i are obtained via thresholding Mi = 1σ(li)>0.52. Applying the full mask M then uncovers the module responsible for the target task. A preliminary study confirmed that the mask training process is stable and thereby suitable for inspection (Appendix B.1).” therefore, an output channel may be created based on the binary mask switching for the first time during the first iteration of training).
The reasons of obviousness have been noted in the rejection of Claim 5 above and applicable herein.
Regarding Claim 8, Courbariaux in view of Helwegen teaches the computing system of claim 1, wherein the processor is configured to:
Courbariaux in view of Helwegen does not explicitly disclose train the RA-BNN using a differentiable Gumbel-Sigmoid method.
However, Csordas teaches train the RA-BNN using a differentiable Gumbel-Sigmoid method (Csordas, Pg. 2, “The binarization is achieved using a Gumbel-Sigmoid with a straight-through estimator, which we derive from the Gumbel-Softmax (Jang et al., 2017; Maddison et al., 2017) in Appendix A.1”, thus, the RA-BNN is trained using a differentiable gumbel-sigmoid method).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the computing system of claim 1, as disclosed by Courbariaux in view of Helwegen to include wherein the RA-BNN is trained using a differentiable Gumbel-Sigmoid method, as disclosed by Csordas. One of ordinary skill in the art would have been motivated to make this modification to enable the use of a gumbel-sigmoid method which allows for differentiable sampling and hence improves feature selection methods of a particular target task (Csordas, Pg. 2, “To train the mask, we treat all N weights separately of each other. Let i [1N] to denote the weight index. The mask’s probabilities are represented as learned logits li R, which are initialized to keep the weights with high probability (0.9). If one were to apply continuous masks to the weights it would be possible to scale them arbitrarily, thereby potentially modifying the function the network performs. To prevent this, we binarize masks, which only provides for keeping or removing individual weights. The binarization is achieved using a Gumbel-Sigmoid with a straight-through estimator, which we derive from the Gumbel-Softmax (Jang et al., 2017; Maddison et al., 2017) in Appendix A.1.”).
Regarding Claim 17, Courbariaux in view of Helwegen teaches the method of claim 16.
Courbariaux in view of Helwegen does not explicitly disclose wherein training and channel-wise growing the BNN from the initial BNN to the larger BNN comprises learning binary masks associated with each weight channel.
However, Csordas teaches wherein training and channel-wise growing the BNN from the initial BNN to the larger BNN comprises learning binary masks associated with each weight channel (Csordas, Pg. 3, “Training is done by applying the loss function defined by the target task and backpropagating into the logits li. Typically multiple (between 4–8) binary masks are sampled and applied to different parts of a batch to improve the quality of the estimated gradient.”, therefore, the training of the network includes learning binary masks associated with each weight channel).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the method for training a BNN of claim 16, as disclosed by Courbariaux in view of Helwegen to include wherein training and channel-wise growing the BNN from the initial BNN to the larger BNN comprises learning binary masks associated with each weight channel, as disclosed by Csordas. One of ordinary skill in the art would have been motivated to make this modification to enable the use of binary weight masks which may identify individual weights responsible for specific functions, hence improving system accuracy and performance for a target function (Csordas, Pg. 1, “In this paper, we present a novel method based on learning binary weight masks to identify individual weights and subnets responsible for specific functions. Using this powerful tool, we contribute an extensive study of emerging modularity in NNs that covers several standard architectures and datasets. We demonstrate how common NNs fail to reuse submodules and offer new insights into the related issue of systematic generalization on language tasks.”)
10. Claim 15 is rejected under 35 U.S.C. 103 as being unpatentable over Courbariaux et al. (hereinafter Courbariaux) (“Binarized Neural Networks: Training Neural Networks with Weights and Activations Constrained to +1 or -1”), in view of Helwegen (hereinafter Helwegen) (US PG-PUB 20220405576), further in view of Wang et al. (hereinafter Wang) (“Learning Channel-wise Interactions for Binary Convolutional Neural Networks”).
Regarding Claim 15, Courbariaux in view of Helwegen teaches the method of claim 13.
Courbariaux in view of Helwegen does not explicitly disclose stopping channel-wise growing the BNN when network growth becomes stable.
However, Wang teaches stopping channel-wise growing the BNN when network growth becomes stable (Wang, Pg. 572, “Finally, we take the action stop to terminate the current epoch of channel-wise interaction mining when the policy network converges or achieves the maximal steps”, therefore, channel-wise growth is stopped when network growth becomes stable/converges).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of claim 13, as disclosed by Courbariaux in view of Helwegen to include stopping channel-wise growing the BNN when network growth becomes stable, as disclosed by Wang. One of ordinary skill in the art would have been motivated to make this modification to enable the stopping/termination of channel-wise interaction/growth which may prevent overfitting and improve overall efficiency once the model has converged (Wang, Pg. 572, “Finally, we take the action stop to terminate the current epoch of channel-wise interaction mining when the policy network converges or achieves the maximal steps”).
11. Claim 18 is rejected under 35 U.S.C. 103 as being unpatentable over Courbariaux et al. (hereinafter Courbariaux) (“Binarized Neural Networks: Training Neural Networks with Weights and Activations Constrained to +1 or -1”), in view of Helwegen (hereinafter Helwegen) (US PG-PUB 20220405576), in view of Csordas et al. (hereinafter Csordas) (“Are Neural Nets Modular? Inspecting Functional Modularity through Differentiable Weight Masks”), further in view of Wang et al. (hereinafter Wang) (“Learning Channel-wise Interactions for Binary Convolutional Neural Networks”).
Regarding Claim 18, Courbariaux in view of Helwegen in view of Csordas teaches the method of claim 17 including when network growth becomes stable: starting retraining the larger BNN to minimize the defined loss (Courbariaux, Pg. 5, Figure 1, which depicts the training/retraining of the BNN across different data sets and minimizes a defined loss/square hinge loss).
Courbariaux in view of Helwegen in view of Csordas does not explicitly disclose when network growth becomes stable: stopping channel-wise growing the BNN;
However, Wang teaches when network growth becomes stable: stopping channel-wise growing the BNN (Wang, Pg. 572, “Finally, we take the action stop to terminate the current epoch of channel-wise interaction mining when the policy network converges or achieves the maximal steps”, therefore, channel-wise growth is stopped when network growth becomes stable/converges);
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the method of claim 17, as disclosed by Courbariaux in view of Helwegen in view of Csordas to include stopping channel-wise growing the BNN when network growth becomes stable, as disclosed by Wang. One of ordinary skill in the art would have been motivated to make this modification to enable the stopping/termination of channel-wise interaction/growth which may prevent overfitting and improve overall efficiency once the model has converged (Wang, Pg. 572, “Finally, we take the action stop to terminate the current epoch of channel-wise interaction mining when the policy network converges or achieves the maximal steps”).
Conclusion
12. 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.
13. Any inquiry concerning this communication or earlier communications from the examiner should be directed to Devika S Maharaj whose telephone number is (571)272-0829. The examiner can normally be reached Monday - Thursday 8:30am - 5:30pm.
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, Alexey Shmatov can be reached at (571)270-3428. 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.
/DEVIKA S MAHARAJ/Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights