DETAILED ACTION
This action is responsive to the Claims filed on 11/14/2023. Claims 1-17 are pending in the case. Claims 1 and 9 are independent claims.
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 .
Priority
Applicant’s claim for the benefit of a prior-filed application under 35 U.S.C. 119(e) or under 35 U.S.C. 120, 121, 365(c), or 386(c) is acknowledged. Applicant has not complied with one or more conditions for receiving the benefit of an earlier filing date under 35 U.S.C. 112(a) as follows:
The disclosure of the prior-filed application, Application No. 16/001,923, fails to provide adequate support or enablement in the manner provided by 35 U.S.C. 112(a) or pre-AIA 35 U.S.C. 112, first paragraph for one or more claims of this application:
Each independent claim, including claim 1, discloses interlayer and intralayer sparsity analysis, visualization of calculated values on a histogram, and pruning a model which is represented by low rank approximation which is then quantized. None of these features are supported in the prior-filed application.
Accordingly, claims 1-17 are not entitled to the benefit of the prior application
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 2-4, 10-12 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
The term “small weight” in the claims are a relative term which renders the claim indefinite. The term “small” is not defined by the claim, the specification does not provide a standard for ascertaining the requisite degree, and one of ordinary skill in the art would not be reasonably apprised of the scope of the invention.
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 1-17 are rejected under 35 U.S.C. 101 because the claim are directed to an abstract idea without significantly more.
Regarding Claim 1/9:
Under step 1, claim 1 is directed to a self-tuning model compression methodology for reconfiguring a Deep Neural Network (DNN) which is directed to a process, one of the statutory categories.
Under step 1, claim 9 is directed to an electronic device which is directed to a machine, one of the statutory categories
Under Step 2A Prong 1, the claim recites the following limitations which are considered mental evaluations:
performing an inter-layer sparsity analysis of the pre-trained DNN model to generate a first sparsity result;
performing an intra-layer sparsity analysis of the pre-trained DNN model to generate a second sparsity result, comprising:
defining a plurality of sparsity metrics for the network;… using the collected data to calculate values for the defined sparsity metrics
according to the first and second sparsity results, performing low-rank approximation on the pre-trained DNN to represent the pre-trained DNN model with low-rank counterparts;
pruning the represented DNN model according to the first and second sparsity results;
performing quantization on the pruned DNN model according to the first and second sparsity results to generate a reconfigured model of the DNN;
Each of these amount to mental evaluation because they describe manipulation of abstract data. Sparsity analysis broadly describes and analysis of data associated with the DNN model, further rank approximation, pruning and quantization are mathematical operations on abstract data which can be performed in the mind.
Under step 2A Prong 2, The claim recites the following additional element(s):
performing forward and backward passes to collect data corresponding to the sparsity metrics; … and executing the reconfigured model on a user terminal for an end-user application.( describes an application at a high degree of generality which makes use of the recited exception, see MPEP 2106.05(f))
receiving a pre-trained DNN model and a data set… and visualizing the calculated values using at least a histogram; (that amounts to adding insignificant extra-solution activity to the judicial exception, because the limitation describe mere data gathering. See MPEP 2106.05(g)
wherein the pre-trained DNN model comprises an input layer, at least one hidden layer and an output layer, and said at least one hidden layer and the output layer of the pre-trained DNN model comprise a plurality of neurons; (is generally linking the use of the judicial exception to a particular technological environment or field of use. The limitation is merely an “incidental or token addition to the claim that did not alter or affect how” the claimed steps are performed, see MPEP 2106.05(h))
Therefore the claim is directed to a judicial exception.
Under step 2B:
receiving a pre-trained DNN model and a data set is well understood, routine, and conventional activity because it amounts to “transmitting or receiving data over a network" (see MPEP 2106.05(d)(II)(i)
The additional elements of visualizing the calculated values using at least a histogram are insignificant extra-solution activities that are considered well-understood, routine, conventional activities. In accordance with the MPEP, the following factual determination is based on the technical publication: [Brasnett et al. US Document ID US 20190354844 A1 (PTO-892)]. Paragraph 0077
“Another common operation in traditional computer vision algorithms is a histogram operation. As shown in FIG. 8 a histogram operation involves dividing the range of values in the input image 802 into intervals called bins and counting how many values fall within each bin to generate a histogram”
discloses that generation of histograms is traditional practice in computer algorithms and common (corresponds to routine and conventional) generation of a histogram given a range of values (corresponds to visualizing the calculated values using at least a histogram). As such, the insignificant extra-solution activities are considered well-understood, routine, conventional activities.
Under step 2B, the recited additional elements when considered alone or in combination neither integrates the abstract idea into a practical application nor provides significantly more than the abstract idea itself.
Regarding Claim 2/10
The claim depends on the rejected parent claim.
Each of the limitations described in the claim, under Step 2A Prong 1, only serve to describe the abstract ideas addressed in the independent claim, in particular the limitations describe mental evaluations.
Furthermore, under step 2A Prong 2 and 2B, the claim does not recite additional elements to consider.
Regarding Claim 3/11
The claim depends on the rejected parent claim.
Under Step 2A Prong 1, The claim recites the limitations:
the collected data comprises weight data obtained by extracting weights for all channels in the pre-trained DNN model, and activation data obtained by monitoring activations for each channel in the pre-trained DNN model after performing a forward pass.
which further describe the abstract ideas recited in the parent claims, under Step 2A Prong 1, in particular the limitations describe mental evaluations. Monitoring respective data amounts to observation of abstract data.
Furthermore, under step 2A Prong 2 and 2B, the claim does not recite additional elements to consider.
Regarding Claim 4/12
The claim depends on the rejected parent claim.
Under Step 2A Prong 1, The claim recites the limitations:
wherein the weight data is used to calculate the percentage of zeroes and the percentage of small weights, and the activation data is used to calculate an average activation value or a percentage of activations below a certain threshold to compute an L1 norm for each channel.
which further describe the abstract ideas recited in the parent claims, under Step 2A Prong 1, in particular the limitations describe mental evaluations. Calculation of a percentage or average is a mental evaluation made in the mind, separately it can also be considered a mathematical operation.
Furthermore, under step 2A Prong 2 and 2B, the claim does not recite additional elements to consider.
Regarding Claim 5/15
The claim depends on the rejected parent claim.
The claim does not recite further abstract idea to consider, beyond those recited in the parent claim.
Under step 2A Prong 2, The claim recites the following further additional element(s):
wherein the DNN model is used for computer vision targeted application models including an AlexNet, a VGG16, a ResNet, and a MobileNet, and natural language understanding application models (is generally linking the use of the judicial exception to a particular technological environment or field of use. The limitation is merely an “incidental or token addition to the claim that did not alter or affect how” the claimed steps are performed, see MPEP 2106.05(h))
Under step 2B, the recited additional elements when considered alone or in combination neither integrates the abstract idea into a practical application nor provides significantly more than the abstract idea itself.
Regarding Claim 6/16
The claim depends on the rejected parent claim.
The claim does not recite further abstract idea to consider, beyond those recited in the parent claim.
Under step 2A Prong 2, The claim recites the following further additional element(s):
wherein each of said at least one hidden layer and the output layer of the reconfigured model is a convolutional layer or a fully-connected layer. (is generally linking the use of the judicial exception to a particular technological environment or field of use. The limitation is merely an “incidental or token addition to the claim that did not alter or affect how” the claimed steps are performed, see MPEP 2106.05(h))
Under step 2B, the recited additional elements when considered alone or in combination neither integrates the abstract idea into a practical application nor provides significantly more than the abstract idea itself.
Regarding Claim 7/17
The claim depends on the rejected parent claim.
The claim does not recite further abstract idea to consider, beyond those recited in the parent claim.
Under step 2A Prong 2, The claim recites the following further additional element(s):
wherein the end-user application is a visual recognition application or a speech recognition application (is generally linking the use of the judicial exception to a particular technological environment or field of use. The limitation is merely an “incidental or token addition to the claim that did not alter or affect how” the claimed steps are performed, see MPEP 2106.05(h))
Under step 2B, the recited additional elements when considered alone or in combination neither integrates the abstract idea into a practical application nor provides significantly more than the abstract idea itself.
Regarding Claim 8/13
The claim depends on the rejected parent claim.
Each of the limitations described in the claim, under Step 2A Prong 1, only serve to describe the abstract ideas addressed in the independent claim, in particular the limitations describe mental evaluations.
Furthermore, under step 2A Prong 2 and 2B, the claim does not recite additional elements to consider.
Regarding Claim 14
The claim depends on the rejected parent claim.
Under Step 2A Prong 1, The claim recites the limitations:
and the DNN model is compressed by removing a portion of the multiplexers and adders in the DNN model according to the analysis result so that a number of multiplexers and adders in the reconfigured model is less than a number of multiplexers and adders in the DNN model.
which further describe the abstract ideas recited in the parent claims, under Step 2A Prong 1, in particular the limitations describe mental evaluations. Removing a portion of logic elements amounts to a mental decision to ignore a set of elements, which is a decision practically made in the human mind.
wherein each of the plurality of neurons of the reconfigured model corresponds to at least one of a multiplexer and an adder, each of the plurality of neurons of the DNN model corresponds to at least one of a multiplexer and an adder, (is generally linking the use of the judicial exception to a particular technological environment or field of use. The limitation is merely an “incidental or token addition to the claim that did not alter or affect how” the claimed steps are performed, see MPEP 2106.05(h))
Under step 2B, the recited additional elements when considered alone or in combination neither integrates the abstract idea into a practical application nor provides significantly more than the abstract idea itself.
Claim Rejections - 35 U.S.C. § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. §§ 102 and 103 (or as subject to pre-AIA 35 U.S.C. §§ 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. § 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102 of this title, 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.
Claim(s) 1, 6-8, 9, 13, and 16-17 are rejected under 35 U.S.C. § 103 as being unpatentable over Han “Learning both Weights and Connections for Efficient Neural Networks” further in view Li “LoSparse: Structured Compression of Large Language Models based on Low-Rank and Sparse Approximation”, further in view of Song Han et al. “deep compression: compressing deep neural networks with pruning, trained quantization and huffman coding” hereinafter Han2
Regarding claim 1/9,
Han teaches, A self-tuning model compression methodology for reconfiguring a Deep Neural Network (DNN), comprising: (abstract pg 1 “Our method prunes redundant connections using a three-step method. First, we train the network to learn which connections are important. Next, we prune the unimportant connections. Finally, we retrain the network to fine tune the weights of the remaining connections. On the ImageNet dataset, our method reduced the number of parameters of AlexNet” Alexnet is a deep neural network) [claim 9] An electronic device, comprising: a storage device, arranged to store a program code; and a processor, arranged to execute the program code; wherein when loaded and executed by the processor, the program code instructs the processor to execute the following steps (pg 4 Section 4 “We implemented network pruning in Caffe…We carried out the experiments on Nvidia TitanX and GTX980 GPUs”) receiving a pre-trained DNN model and a data set, (Figure 2 and Section 3 pg 3 “Our pruning method employs a three-step process, as illustrated in Figure 2, which begins by learning the connectivity via normal network training…The second step is to prune the low-weight connections” the first step is to train a model, the resulting model to be pruned and analyzed is thus a pre-trained model Section 4.1 pg 4 “We first experimented on MNIST dataset”) wherein the pre-trained DNN model comprises an input layer, at least one hidden layer and an output layer, and said at least one hidden layer and the output layer of the pre-trained DNN model comprise a plurality of neurons; (Figure 3 Section 3 pg 3 “All connections with weights below a threshold are removed from the network — converting a dense network into a sparse network, as shown in Figure 3.
PNG
media_image1.png
273
438
media_image1.png
Greyscale
the neural network in question is deep as it has many layers comprising at least in the figure 2 hidden layers, and input layer and output layer. Further the system is tested according to many different named deep neural networks, See Section 4) performing an inter-layer sparsity analysis of the pre-trained DNN model to generate a first sparsity result; performing an intra-layer sparsity analysis of the pre-trained DNN model to generate a second sparsity result, (pg 6 Table 4 “Table4: For AlexNet, pruning reduces the number of weights by 9× and computation by 3×.”
PNG
media_image2.png
247
822
media_image2.png
Greyscale
interlayer sparsity is the percentage of activations in layers, while intra-layer sparsity can be considered the percentage of weights. Examiner notes the disclosure provides no definition for inter or intra layer sparsity. The broadest reasonable interpretation is merely sparsity within layers and sparsity between layers.) comprising: defining a plurality of sparsity metrics for the network; (pg 6 Table 4 “Table4: For AlexNet, pruning reduces the number of weights by 9× and computation by 3×.” The table defines the sparsity metrics including the percentage of activation weights and flops.) performing forward and backward passes to collect data corresponding to the sparsity metrics; using the collected data to calculate values for the defined sparsity metrics; and visualizing the calculated values using at least a histogram; (pg 6 Table 4 “Table4: For AlexNet, pruning reduces the number of weights by 9× and computation by 3×.”
PNG
media_image2.png
247
822
media_image2.png
Greyscale
pg 8 Figure 7 “Figure7: Weight distribution before and after parameter pruning. The right figure has 10× smaller scale.” The distribution of weights is a histogram of calculated weights from training which involves the claimed forward and backward passes
PNG
media_image3.png
224
854
media_image3.png
Greyscale
)
Han does not explicitly teach, according to the first and second sparsity results, performing low-rank approximation on the pre-trained DNN to represent the pre-trained DNN model with low-rank counterparts… pruning the represented DNN model according to the first and second sparsity results;… performing quantization on the pruned DNN model according to the first and second sparsity results to generate a reconfigured model of the DNN… and executing the reconfigured model on a user terminal for an end-user application.
Li however when addressing improvements to pruning using low rank approximations teaches, according to the first and second sparsity results, performing low-rank approximation on the pre-trained DNN to represent the pre-trained DNN model with low-rank counterparts ( pg 3 Section 3 “We propose a compression method for transformer models. Specifically, we approximate a weight matrix by the sum of a low-rank matrix and a sparse matrix…” pg 4 Section 3.2 “We then present our proposed algorithm. Given a pre-trained weight matrix W(0), we first initialize the low-rank matrix of rank r based on the singular value decomposition (SVD) of W(0)… We apply such a decomposition to every weight matrix of the model and denote S={Sm}M m=1 as the set of all sparse matrices” the weight matrix of the pre-trained DNN is subdivided using SVD which is a set of low rank counterparts for the connection matrix of the neural network. The connections are according to a given sparsity by the matrix S in the decomposition) pruning the represented DNN model according to the first and second sparsity results; (pg 4 Section 3.2 “After the initialization, we conduct the iterative structured pruning for S” S is the set of sparse matrices thus includes first and second sparsity results)
Accordingly, it would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify pruning neural network method described by Han to comprise the improved pruning which includes low rank approximation of the neural network described by Li. One would have been motivated to make such a combination because Han and Li describe deep neural network pruning and of pre trained models. Further, Li notes that “the low-rank approximation prevents the pruning from excessively removing expressive neurons while sparse approximation enhances the diversity of low-rank approximation” (abstract Li). Further Li notes, “our method significantly surpasses previous compression approaches” (Li Conclusion)
Han/Li does not explicitly teach, performing quantization on the pruned DNN model according to the first and second sparsity results to generate a reconfigured model of the DNN; and executing the reconfigured model on a user terminal for an end-user application.
Han2 however teaches, performing quantization on the pruned DNN model according to the first and second sparsity results to generate a reconfigured model of the DNN; (pg 2 Figure 1 caption “The three stage compression pipeline: pruning, quantization and Huffman coding. Pruning reduces the number of weights by 10×, while quantization further improves the compression rate” the quantization occurs after pruning thus according to the determined sparsity. The resulting model is the reconfigured DNN) and executing the reconfigured model on a user terminal for an end-user application. (pg 5 Section 5 “We pruned, quantized, and Huffman encoded four networks: two on MNIST and two on ImageNet data-sets.” The system is employed on two data sets Figure 6 pg 8 “Accuracy v.s. compression rate under different compression methods” the results are depicted which came from a user terminal for end use applications)
Accordingly, it would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify low rank approximation and pruning neural network method described by Han/Li to comprise the quantization of pruned neural networks described by Han2. One would have been motivated to make such a combination because Han and Li and Han2 describe deep neural network pruning and of pre trained models. Further, Han2 notes “Our main insight is that pruning and trained quantization are able to compress the network without interfering each other, thus lead to surprisingly high compression rate. It makes the required storage so small” (Han2 pg 2)
Regarding claim 6/16
Han/Li/Han2 teaches parent claim 1/9
Further Han teaches, wherein each of said at least one hidden layer and the output layer of the reconfigured model is a convolutional layer or a fully-connected layer. (pg 5 Section 4.2 “We use the AlexNet Caffe model as the reference model, which has 61 million parameters across 5 convolutional layers and 3 fully connected layers” pg 3 Figure 3 “Figure 3: Synapses and neurons before and after pruning.”
PNG
media_image4.png
202
430
media_image4.png
Greyscale
the figure clearly depicts the original pre-pruned network as fully connected for both hidden and output layers.)
Regarding claim 7/17
Han/Li/Han2 teaches parent claim 1/9
Further Han teaches, wherein the end-user application is a visual recognition application or a speech recognition application. (pg 8 Section 6 “We highlight our experiments on AlexNet and VGGNet on ImageNet, showing that both fully connected layer and convolutional layer can be pruned, reducing the number of connections by 9× to 13× without loss of accuracy. …for real-time image processing, making it easier to be deployed on mobile systems.” The ImageNet tasks are a visual recognition application)
Regarding claim 8/13
Han/Li/Han2 teaches parent claim 1/9
Further Han teaches, wherein a number of the plurality of neurons of the reconfigured model is less than a number of the plurality of neurons of the DNN model. (pg 3 Figure 3 “Figure 3: Synapses and neurons before and after pruning.”
PNG
media_image4.png
202
430
media_image4.png
Greyscale
the figure clearly depicts the original pre-pruned network as fully connected for both hidden and output layers.)
Claim(s) 2-4 and 10-12 are rejected under 35 U.S.C. § 103 as being unpatentable over Han/Li/Han2, further in view of Zhang et al “Learning Low-Precision Structured Subnetworks Using Joint Layerwise Channel Pruning and Uniform Quantization”
Regarding claim 2/10
Han/Li/Han2 teaches parent claim 1/9
Further Han teaches, wherein the defined sparsity metrics comprise percentage of zeroes, small weight percentage (pg 8 Figure 7 “Figure7: Weight distribution before and after parameter pruning. The right figure has 10× smaller scale.” The distribution of weight values depicts the percentage or proportion of small and zero weights.
PNG
media_image3.png
224
854
media_image3.png
Greyscale
pg 5 “the average percentage of activations that are non-zero, the percentage of non-zero weights after pruning, and the percentage of actually required floating point operations.” Table 2 pg 5 “
PNG
media_image5.png
113
546
media_image5.png
Greyscale
the surviving weight percentage thus indicates the percentages of zero or small weights.)
Han/Li/Han2 does not explicitly teach, wherein the defined sparsity metrics… and L1 norms
Zhang however when addressing sparsity metrics such as L1 norms teaches, wherein the defined sparsity metrics… and L1 norms (pg 6 Section 3 “To rank order output channels by importance, we use two measures: (1) the mean ℓ1-norm of the output activations generated by each channel, and (2) the ℓ1-norm of the learned weights for each output channel.” Pg 7 “We increase the sparsity of the input channels using the rank ordering of the original unpruned network as determined by the mean ℓ1-norm of its activations when estimated over the test dataset.”)
Accordingly, it would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify low rank approximation and pruning neural network method described by Han/Li/Han2 to comprise the normalization based pruning described by Zhang. One would have been motivated to make such a combination because Han/Li/Han2 and Zang describe deep neural network pruning and of pre trained models. Further, Zhang notes “we demonstrate increased performance per memory footprint over existing solutions across a range of discriminative and generative networks.” (abstract Zhang)
Regarding claim 3/11
Han/Li/Han2/Zhang teaches parent claim 2/10
Han teaches, wherein the collected data comprises weight data obtained by extracting weights for all channels in the pre-trained DNN model and activation data obtained by monitoring activations for each channel in the pre-trained DNN model after performing a forward pass. (pg 5 Table 2… “For Lenet-300-100, pruning reduces the number of weights”
PNG
media_image6.png
130
557
media_image6.png
Greyscale
the presented data shows the percentage of weights and the activation data (act%) this is collected as a result of the forward pass required for the pruning described and claimed.)
Regarding claim 4/12
Han/Li/Han2/Zhang teaches parent claim 3/11
Han teaches, wherein the weight data is used to calculate the percentage of zeroes and the percentage of small weights, and the activation data is used to calculate an average activation value or a percentage of activations below a certain threshold. (pg 5 Table 2… “For Lenet-300-100, pruning reduces the number of weights”
PNG
media_image6.png
130
557
media_image6.png
Greyscale
pg 5 “For each layer of the network the table shows (left to right) the original number of weights, the number of floating point operations to compute that layer’s activations, the average percentage of activations that are non-zero,” the weights are a percentage of non-zero weights thus are a calculation of the percentage of zero and small weights, the activations are the average percentage of activation that are non-zero thus a calculation of percentage below a threshold (i.e the percentage of activation above the threshold reflects the percentage below the threshold as they are inverses)
Zhang teaches, to compute an L1 norm for each channel. (pg 6 Section 3 “To rank order output channels by importance, we use two measures: (1) the mean ℓ1-norm of the output activations generated by each channel, and (2) the ℓ1-norm of the learned weights for each output channel.” Pg 7 “We increase the sparsity of the input channels using the rank ordering of the original unpruned network as determined by the mean ℓ1-norm of its activations when estimated over the test dataset.” The determined activations and weights of the network serve to compute the output channel mean ℓ1-norm )
Claim(s) 5 and 15 are rejected under 35 U.S.C. § 103 as being unpatentable over Han/Li/Han2, further in view of Gadosey “On Pruned, Quantized and Compact CNN Architectures for Vision Applications”
Regarding claim 5/15
Han/Li/Han2 teaches parent claim 1/9
Han teaches, wherein the DNN model is used for computer vision targeted application models including an AlexNet, a VGG16 (pg 4 Section 4 “We pruned four representative networks: Lenet-300-100 and Lenet-5 on MNIST, together with AlexNet and VGG-16 on ImageNet.” ImageNet is a visual objection recognition application, thus computer vision)
Li teaches, and natural language understanding application models. (pg 2 “We conduct extensive experiments on natural language understanding, question answering, and natural language generation tasks to demonstrate the effectiveness and efficiency of LoSparse.”)
Gadosey when addressing experiments with several types of models teaches, a ResNet, and a MobileNet (abstract pg 1 “The authors of this paper review and experiment with compact models (MobileNet V1 and V2, ShuffleNet V1 and V2, FD MobileNet)) and selected methods of pruning and quantization” pg 4 Section 4.2 “The proposed method was first tested with the VGG-16, VGG-19 and ResNet18 networks”)
Accordingly, it would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify low rank approximation and pruning neural network method described by Han/Li/Han2 to comprise the pruning methods for a variety of models as described by Gadosey. One would have been motivated to make such a combination because Han/Li/Han2 and Gadosey describe deep neural network pruning and of pre trained models. Further, Gadosey notes “reduces computational requirements significantly with improvement in accuracy over baseline models” (Gadosey pg 2)
Claim(s) 4 are rejected under 35 U.S.C. § 103 as being unpatentable over Han/Li/Han2, further in view of Saffar “A Neural Network Architecture Using High Resolution Multiplying Digital to Analog Converters”
Regarding claim 14
Han/Li/Han2 teaches parent claim 9
Han/Li/Han2 does not teach, wherein each of the plurality of neurons of the reconfigured model corresponds to at least one logic circuit comprising at least one of a multiplexer and an adder, each of the plurality of neurons of the DNN model corresponds to at least one logic circuit comprising at least one of a multiplexer and an adder, and compressing the DNN model into a reconfigured model comprises; removing a portion of the logic circuits in the DNN model according to the analysis result so that a number of logic circuits in the reconfigured model is less than a number of logic circuits in the DNN model.
Saffar explicitly discloses the logic circuit and multiplexer or adder hardware realization of a neural network teaching, wherein each of the plurality of neurons of the reconfigured model corresponds to at least one of a multiplexer and an adder, each of the plurality of neurons of the DNN model corresponds to at least one of a multiplexer and an adder, and the DNN model is compressed (Saffar Section 2 “Fig. 1 shows the block diagram of one layer of the proposed network… Figure 1, where m and n represent the number of corresponding neuron and inputs to each layer, respectively….The number of MDACs in each layer is equal to the number of inputs to that layer… The adder block is simply a current-mode summation node…There is one time-multiplexer per each layer, which acts as the selection signal controlling inputs to the MDAC block…” the network consisting of a plurality of neurons and layers, composed of multiplexer and adders.)
Han/Li/Han2 when combined with Saffar teaches, and the DNN model is compressed by removing a portion of the multiplexers and adders in the DNN model according to the analysis result so that a number of multiplexers and adders in the reconfigured model is less than a number of multiplexers and adders in the DNN model (pg 3 Figure 3 “Figure 3: Synapses and neurons before and after pruning.”
PNG
media_image4.png
202
430
media_image4.png
Greyscale
, after pruning model has less neuron than before pruning model. Figures 1-3 of Saffar depict many circuit components related to the neural network, when a neuron is removed the number of circuit components necessarily is reduced.)
Han/Li/Han2 and Saffar both disclose neural network acceleration application and are analogous. It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combining Han/Li/Han2 teaching of neural network compression method with Saffar’s teaching of acceleration apparatus to achieve the claimed teaching. One of the ordinary skilled in the art would have motivated to make this modification in order to reduce circuit size (Conclusion Saffar “can significantly reduce the circuit size and makes the proposed method feasible to be implemented in mixed-signal neural network”)
Conclusion
prior art not relied upon:
Guan Li et al. “CNNPruner: Pruning Convolutional Neural Networks with Visual Analytics” describes an interactive user guided pruning method with user terminal.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JOHNATHAN R GERMICK whose telephone number is (571)272-8363. The examiner can normally be reached M-F 9:30-4:30.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Kakali Chaki can be reached on 571-272-3719. 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.
/J.R.G./
Examiner, Art Unit 2122
/KAKALI CHAKI/Supervisory Patent Examiner, Art Unit 2122
Prosecution Insights