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 .
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 21-28, 30-38, 40 are rejected under 35 U.S.C. 103 as being unpatentable over Sun et.al (NPL: Weights Having Stable Signs Are Important: Finding Primary Subnetworks and Kernels to Compress Binary Weight Networks) in view of Lin et.al (US 20230086378 A1).
Regarding claim 21
Sun teaches or at least suggests “for a convolutional layer of the CNN model: allocate a 1-bit convolutional kernel subset to the convolutional layer, wherein the convolutional layer includes 32-bit or 16-bit floating-point convolutional kernels with a size of K x K and the 1-bit convolutional kernel subset includes 2N 1-bit convolutional kernel candidates with the size of K x K, 1 ≤ N < K x K and both K and N are positive integers” (Page 1 “In the optimization, they learn optimal binary weight outputs represented as a combination of scaling factors and weight signs to approximate 32-bit floating-point weight values, usually with a layer-wise quantization scheme”, Page 5 section 2.6 “For a binary-kernel with 3×3 kernel size, there are 29 possible kernels in total”, Page 6 section 3.1 “we first train an ordinary VGG-7 XNor-BWN on Cifar-10 and extract its last Conv layer’s binary kernel distribution ... Then we sort these binary kernels according to their appearance frequency and select top 21, 22, ..., 28 frequent binary kernels. These kernels are called selected binary-kernels ... we use the selected binary kernels to indicate the kernels”, and Page 7 section 3.1-3.2 “we use the quantized bit numbers for each layer to indicate how many selected quantized kernels are used ... When we use the quantized bit p < 9, we can use less than 9-bit number to represent the binary-kernel, this provides the compression ability of QBN”. Sun discloses a binary-weight CNN in which full-precision weights are retained during training and binary weights are used to approximate 32-bit floating-point weight values. For the convolutional layers, Sun discloses binary kernels having a 3×3 kernel size, with 29 possible binary-kernel patterns. Sun further sorts the binary kernels according to their appearance frequency and selects the top 21, 22, ..., 28 frequent binary kernels, referred to as “selected binary-kernels” Km. The 2p selected binary kernels correspond to the claimed 1-bit convolutional kernel subset, with each selected binary kernel Km corresponding to a 1-bit convolutional kernel candidate, because the selected binary kernels define the limited collection of binary-kernel patterns available for subsequent quantization. Sun further discloses using a quantized bit number p for each layer to indicate how many selected quantized kernels are used by that layer. Thus, Sun’s use of 2p selected binary kernels for a respective convolutional layer teaches or suggests allocating the corresponding 1-bit convolutional kernel subset to that convolutional layer. Further, because Sun discloses 3×3 binary kernels, K=3 and K x K = 9; and because Sun selects 2p binary kernels with p=1, ... 8, Sun’s p corresponds to claimed N, such that the subset includes 2N 1-bit convolutional kernel candidates and satisfies 1 ≤ N < K x K.)
Sun teaches or at least suggests “perform weights quantization of the convolutional layer by selecting 1-bit convolutional kernel candidates from the 1-bit convolutional kernel subset as 1-bit convolutional kernels of the convolutional layer” (Page 6 section 3.1 “we sort these binary kernels according to their appearance frequency and select top 21, 22, ..., 28 frequent binary kernels. These kernels are called selected binary-kernels ... we use the selected binary kernels to indicate the kernels”, and Page 7 section 3.1 algorithm 1“FP indicates the first full-precision Conv layer which is not quantized according to the common practice ... we use the quantized bit numbers for each layer to indicate how many selected quantized kernels are used ... the full-precision kernel will be replaced by the selected kernel whose distance to the full-precision kernel is the shortest one during forward ... Algorithm 1 ... for m in range (2p) do
L2(m)=||Wij −Km||2
end
m∗ = argminm(L2(m))
Wij = αKm∗
...” Sun discloses quantizing a full-precision convolutional kernel by selecting one binary kernel from the previously selected collection of binary kernels. In particular, Sun’s Algorithm 1 considers the 2^p selected binary kernels Km, calculates the L2 distance between a full-precision kernel Wij and each selected binary kernel Km, and determines m* corresponding to the selected binary kernel having the minimum distance. Sun further explains that the full-precision kernel is replaced by the selected kernel whose distance to the full-precision kernel is the shortest during forward. Accordingly, each Km represents one possible binary-kernel candidate within the 2p selected binary-kernel subset, and Sun’s determination of Wij selects one such candidate from that subset for quantizing the full-precision convolutional kernel. Because the weights of the convolutional layer are contained in its full-precision convolutional kernels Wij replacing each such full-precision kernel with a selected binary kernel Km converts the corresponding layer weights from full precision to binary form and thereby performs weight quantization of the convolutional layer. Thus, Sun teaches or suggests performing weights quantization of the convolutional layer by selecting a 1-bit convolutional kernel candidate from the 1-bit convolutional kernel subset as a 1-bit convolutional kernel of the convolutional layer, as claimed.)
Sun does not teach the following limitations “memory;”, “computer executable instructions;”, and “processor circuitry to be programmed by the computer executable instructions”. However, Lin teaches these limitations (paragraph 141 “A system, comprising: a memory comprising computer-executable instructions; and one or more processors configured to execute the computer-executable instructions and cause the processing system to perform a method”. Lin discloses a system comprising a memory having computer-executable instructions and one or more processors configured to execute the computer-executable instructions to cause the processing system to perform a method. Thus, Lin teaches the claimed memory, computer-executable instructions, and processor circuitry programmed by the computer-executable instructions. A person of ordinary skill in the art would have understood that Sun’s disclosed binary-kernel quantization method could be implemented using Lin’s disclosed computer architecture by storing instructions for performing Sun’s kernel-selection and quantization operations in the memory and executing those instructions by the processor, thereby implementing Sun’s CNN quantization method using the claimed memory, computer-executable instructions, and processor circuitry)
Before the effective filing date, it would have been obvious to a person ordinary skill in the past to combine teaching of a binary kernel quantization method that uses a selected set of binary kernels to quantize convolutional-layer kernels and compress binary-weight networks by Sun, with the teaching of refining the weights of each kernel in a convolutional neural network and implementing the convolutional neural network system and method using memory and computer processor with executable program instruction by Lin. The motivation to do so is referred to in Lin’s disclosure (paragraph 74 “a convolutional neural network layer may be associated with one or more kernels, and the weights of each kernel can be iteratively refined based on training data. This allows the kernels to adapt and learn to identify relevant features for the desired output. In at least one aspect, the training can occur incrementally or intermittently during inferencing”, and paragraph 75 “Generally, training the model requires iteratively refining each weight of each kernel ... In embodiments, therefore, use of shaped kernels can accelerate the training procedure by eliminating some kernel elements, and thereby reducing the number of operations that must be performed to update the kernel”. Lin discloses convolutional-kernel weights may be iteratively refined based on training data so that the kernels can adapt and learn to identify relevant features for the desired output and further explains that its kernel technique can accelerate the training procedure by eliminating some kernel elements and thereby reducing the number of operations required to update the kernel. Thus, a person of ordinary skill in the art would have been motivated to implement Sun’s quantization operations using Lin’s memory and processor as a known computer implementation for carrying out CNN processing, and to apply Lin’s iterative refinement to Sun’s selected binary kernels to permit the kernels to adapt during training while improving the efficiency of the kernel-update/training process, with predictable results.)
Regarding claim 22 depends on claim 21, thus the rejection of claim 21 is incorporated.
Sun teaches or at least suggest the limitation “The apparatus of claim 21, wherein the processor circuitry is to update 32-bit or 16- bit floating-point convolutional kernels of respective convolutional layers of the CNN model” (Page 1 “In the optimization, they learn optimal binary weight outputs represented as a combination of scaling factors and weight signs to approximate 32-bit floating-point weight values, usually with a layer-wise quantization scheme” Sun discloses that, in STE-based binary-weight-network training, full-precision weights are retained during training, quantized to binary values during the forward pass, and updated during the backward pass using gradients with respect to the full-precision weights. Sun further explains that the binary weights approximate 32-bit floating-point weight values on a layer-wise basis and identifies Wij as a full-precision convolutional kernel. Thus, updating the full-precision weights forming Wij during training corresponds to updating 32-bit floating-point convolutional kernels of the respective convolutional layers, as claimed.)
Regarding claim 23 depends on claim 22, thus the rejection of claim 22 is incorporated.
Lin teaches or at least suggest the limitation “when a number of training iterations for the CNN model does not reach a preset iteration number, for the convolutional layer of the CNN model” (paragraph 83 “If the training system determines, at block 530, that no additional elements in the shaped kernel remain to be refined, the method 500 continues to block 535 where the training system determines whether training is complete. This may include, for example, ..., determining whether a predefined number of training iterations have been performed” Lin discloses a CNN training process in which the training system determines whether training is complete by determining, among other criteria, whether a predefined number of training iterations has been performed. Thus, before the predefined number of training iterations has been reached, the training-completion criterion has not yet been satisfied, and training continues. Lin therefore teaches or at least suggests the claimed condition in which the number of training iterations for the CNN model has not reached a preset iteration number.)
Sun in view of Lin teaches or at least suggest the limitation “refine the 1-bit convolutional kernel subset, and perform weights quantization of the convolutional layer by selecting 1-bit convolutional kernel candidates from the refined 1-bit convolutional kernel subset as the 1-bit convolutional kernels of the convolutional layer” (Lin discloses at paragraph 74 “a convolutional neural network layer may be associated with one or more kernels, and the weights of each kernel can be iteratively refined based on training data. This allows the kernels to adapt and learn to identify relevant features for the desired output”, and paragraph 78 “Generally, the loss reflects the difference between the actual output and the desired or target output. In some embodiments, this loss can be used to refine one or more model parameters (e.g., weights and biases) in order to improve its accuracy” Lin discloses that a convolutional neural-network layer may include one or more kernels whose weights are iteratively refined based on training data, and further discloses using the computed loss to refine model parameters, including weights, during training. Thus, Lin teaches iterative refinement of convolutional-kernel parameters as CNN training proceeds. Sun further discloses a selected collection of 1-bit convolutional kernels Km that forms the binary-kernel candidate subset and selecting a kernel candidate from that subset to quantize a full-precision convolutional kernel. A person of ordinary skill in the art would have been motivated to apply Lin’s iterative kernel-refinement technique to Sun’s selected binary-kernel candidates during training and thereafter perform Sun’s disclosed quantization by selecting a candidate from the resulting refined binary-kernel subset, thereby teaching or suggesting refinement of the 1-bit convolutional kernel subset followed by weight quantization using the refined subset.)
Sun in view of Lin teaches or at least suggest the limitation “the refined 1-bit convolutional kernel subset includes 2N 1-bit convolutional kernel candidates with the size of K x K” (Sun discloses at page 5 section 2.6 “For a binary-kernel with 3×3 kernel size, there are 29 possible kernels in total”, and at page 6 section 3.1 “we first train an ordinary VGG-7 XNor-BWN on Cifar-10 and extract its last Conv layer’s binary kernel distribution ... Then we sort these binary kernels according to their appearance frequency and select top 21, 22, ..., 28 frequent binary kernels. These kernels are called selected binary-kernels ... we use the selected binary kernels to indicate the kernels” Sun discloses that its selected binary-kernel subset contains 2p binary convolutional-kernel candidates having a 3 x 3 kernel size, as previously discussed with respect to claim 21. Lin’s refinement is directed to refining the kernel weights or parameters during training and does not require changing either the number of kernels in the selected collection or their kernel dimensions. Accordingly, applying Lin’s refinement technique to Sun’s selected binary-kernel subset would retain Sun’s disclosed 2p candidate structure and kernel size, thereby teaching or suggesting a refined subset containing 2N 1-bit convolutional kernel candidates having a size of K x K, as claimed.)
Regarding claim 24 depends on claim 21, thus the rejection of claim 21 is incorporated.
Sun teaches or at least suggests the limitation “The apparatus of claim 21, wherein the 1-bit convolutional kernel subset is shared to all convolutional layers of the CNN model” (Page 6 section 3.1 “we sort these binary kernels according to their appearance frequency and select top 21, 22, ..., 28 frequent binary kernels. These kernels are called selected binary-kernels ... we use the selected binary kernels to indicate the kernels”, Page 7 section 3.1 “FP indicates the first full-precision Conv layer which is not quantized according to the common practice. VGG-7 has 6 Conv layers, and we use the quantized bit numbers for each layer to indicate how many selected quantized kernels are used”, and Page 14 section K “Selected binary kernels do exist, and selected binary-kernels can transfer to other different layers, networks, or datasets” Sun discloses obtaining a common collection of selected binary kernels by sorting binary kernels according to appearance frequency and selecting the top 21, 22, ..., 28 frequent binary kernels. Sun thereafter uses the selected quantized kernels on a layer-by-layer basis and further states that the selected binary kernels can transfer to other different layers, networks, or datasets, thereby teaching reuse of the selected binary-kernel collection across different convolutional layers. Although Sun retains the first convolutional layer at full precision and does not quantize that layer “according to the common practice,” this exclusion is presented as a conventional implementation practice rather than a technical restriction on the QBN technique. It therefore would have been obvious to apply the same selected binary-kernel collection to the first convolutional layer as well, resulting in the 1-bit convolutional kernel subset being shared to all convolutional layers, corresponding to the 1-bit convolutional kernel subset is shared to all convolutional layers of the CNN model, as claimed.)
Regarding claim 25 depends on claim 21, thus the rejection of claim 21 is incorporated.
Sun teaches or at least suggests the limitation “The apparatus of claim 21, wherein the 1-bit convolutional kernel subset is specific to the convolutional layer” (page 7 section 3.1 table 1 “FP indicates the first full-precision Conv layer which is not quantized according to the common practice. VGG-7 has 6 Conv layers, and we use the quantized bit numbers for each layer to indicate how many selected quantized kernels are used”, and Page 8 section 4.3 “When using low quantization bit for binary-kernels, the performance drop will not be negligible, thus how to assign quantization bit to different layer is important” Sun discloses using a quantized bit number for each layer to indicate how many selected quantized kernels are used, and further explains that assigning a quantization bit to different layers is important. Thus, Sun determines the quantized bit number p on a layer-specific basis, with p determining the 2p selected binary kernels used by the respective convolutional layer, thereby teaches or at least suggest that the collection of selected 1-bit kernels is specific to the convolutional layer, which corresponds to the 1-bit convolutional kernel subset is specific to the convolutional layer, as claimed.)
Regarding claim 26 depends on claim 25, thus the rejection of claim 25 is incorporated.
Sun teaches or at least suggests the limitation “The apparatus of claim 25, wherein 1-bit convolutional kernel subsets allocated to different convolutional layers of the CNN model include the same number of 1-bit convolutional kernel candidates” (page 7 section 3.1 table 1 “FP indicates the first full-precision Conv layer which is not quantized according to the common practice. VGG-7 has 6 Conv layers, and we use the quantized bit numbers for each layer to indicate how many selected quantized kernels are used”, and table 1 “VGG-7 ... FP-3-3-3-3-3” Sun discloses that the quantized bit number assigned to each layer indicates how many selected quantized kernels are used by that layer. In Table 1, Sun provides the VGG-7 configuration FP-3-3-3-3-3, in which the different quantized convolutional layers are each assigned the same quantized bit number p=3. Because p determines the 2p selected binary kernels used by a layer, each of those layers uses the same number, 23, of selected binary-kernel candidates, thereby providing the same number of 1-bit selected kernel for the different convolutional layers, which teaches or at least suggest that the 1-bit convolutional kernel subsets allocated to different convolutional layers of the CNN model include the same number of 1-bit convolutional kernel candidates, as claimed.)
Regarding claim 27 depends on claim 25, thus the rejection of claim 25 is incorporated.
Sun teaches or at least suggests the limitation “The apparatus of claim 25, wherein 1-bit convolutional kernel subsets allocated to different convolutional layers of the CNN model include different numbers of 1-bit convolutional kernel candidates” (page 7 section 3.1 table 1 “FP indicates the first full-precision Conv layer which is not quantized according to the common practice. VGG-7 has 6 Conv layers, and we use the quantized bit numbers for each layer to indicate how many selected quantized kernels are used”, and table 1 “VGG-7 ... FP-6-5-4-3-2” Sun discloses the VGG-7 configuration FP-6-5-4-3-2, in which the respective quantized convolutional layers are assigned different quantized bit numbers. Because Sun explains that the quantized bit number for each layer indicates how many selected quantized kernels are used, these respective layers use 26, 25, 24, 23, and 22 selected binary kernels. Thus, different selected binary-kernels are allocated to the different convolutional layers include different numbers of 1-bit convolutional kernel candidates, which teaches or at least suggest that the 1-bit convolutional kernel subsets allocated to different convolutional layers of the CNN model include different numbers of 1-bit convolutional kernel candidates, as claimed.)
Regarding claim 28 depends on claim 21, thus the rejection of claim 21 is incorporated.
Sun teaches or at least suggests the limitation “The apparatus of claim 21, wherein the 1-bit convolutional kernel candidates of the 1-bit convolutional kernel subset are predefined” (Page 6 section 3.1 “we sort these binary kernels according to their appearance frequency and select top 21, 22, ..., 28 frequent binary kernels. These kernels are called selected binary-kernels ... we use the selected binary kernels to indicate the kernels ... In our following experiments these selected binary kernels are extracted from one single VGG-7 BWN’s last Conv layer. After pre-processing these ... we start to train a QBN using Algorithm. 1” Sun discloses that, before training the QBN, it first trains an ordinary VGG-7 XNor-BWN, extracts its binary-kernel distribution, sorts the binary kernels according to appearance frequency, and selects the top 21, 22, ..., 28 frequent binary kernels as the selected binary-kernels Km. Sun then states that, after preprocessing and obtaining these selected binary kernels, it starts training the QBN using Algorithm 1. Thus, the binary-kernel candidates are determined before the QBN training and quantization process in which they are used, corresponding to predefined 1-bit convolutional kernel candidates, as claimed.)
Regarding claim 30 depends on claim 21, thus the rejection of claim 21 is incorporated.
Sun teaches or at least suggests the limitation “The apparatus of claim 21, wherein an objective function of network quantization of the CNN model is defined as follows:
a
r
g
W
t
^
m
i
n
W
t
^
-
W
t
2
2
,
W
t
^
∈
P
,
P
=
{
w
1
,
w
2
,
…
,
w
2
N
}
” (page 6 section 3.1 “we use the selected binary kernels to indicate the kernels K0, 1, ..., 28−1 in our algorithm”, and page 7 Algorithm 1
“Algorithm 1 ...
for m in range (2p) do
L2(m)=||Wij −Km||2
end
m∗ = argminm(L2(m))
Wij = αKm
...
We use L2 norm to calculate the distance between the full-precision kernel Wij to the selected kernels Km, where the full-precision kernel will be replaced by the selected kernel whose distance to the full-precision kernel is the shortest one during forward” Sun discloses that, for each selected binary kernel Km, Algorithm 1 first calculates the L2 distance L2(m)=||Wij − Km||2 from the full-precision kernel Wij, and then determines m∗ = argminm(L2(m)), thereby identifying the selected binary kernel Km* having the shortest distance to Wij. Because m indexes the selected binary kernels K0, 1, ..., 2P−1, Sun’s operation can equivalently be expressed as selecting Km from that collection to minimize ||Wij − Km||2. Thus, Sun’s Km teaches or suggests the claimed
W
t
^
, Sun’s Wij teaches or suggests the claimed Wt, and Sun’s collection of selected binary kernels teaches or suggests P, such that Sun teaches or at least suggest the objective
a
r
g
W
t
^
m
i
n
W
t
^
-
W
t
2
2
. A person of ordinary skill in the art would have found it obvious to modify Sun’s objective to use the squared L2 norm,
W
t
^
-
W
t
2
2
, because L2 distances are nonnegative and squaring them preserves their relative ordering and therefore produces the same minimizing binary kernel, while providing the conventional squared-error form.)
Sun teaches or at least suggests the limitation “wherein
P
=
{
w
1
,
w
2
,
…
,
w
2
N
}
is the 1-bit convolutional kernel subset,
W
t
^
is a quantized weight set of the convolutional layer, and
W
t
is a 32-bit or 16-bit floating point weight set of the convolutional layer” (page 6 section 3.1 Algorithm 1 “we sort these binary kernels according to their appearance frequency and select top 21,22,...,28 frequent binary kernels. These kernels are called selected binary-kernels K0, 1, ..., 28−1 … we use the selected binary kernels to indicate the kernels K0, 1, ..., 28−1”, and page 7 section 3.1 “We use L2 norm to calculate the distance between the full-precision kernel Wij to the selected kernels Km, where the full-precision kernel will be replaced by the selected kernel whose distance to the full-precision kernel is the shortest one during forward” Sun discloses a collection of selected binary kernels K0, 1, ..., 2P−1, a full-precision kernel Wij, and replacing Wij during the forward operation with the selected binary kernel Km* having the shortest distance thereto. Because a convolutional kernel comprises a set of kernel weights, Sun’s full-precision kernel Wij teaches or suggests the claimed floating-point weight set Wt, while Sun’s selected binary kernel Km*, comprising the binary weights used to replace Wij, teaches or suggests the claimed quantized weight set
W
t
^
. Sun’s collection K0, 1, ..., 28−1, from which K Km* is selected, teaches or suggests the claimed P.)
Regarding claim 31,
Lin teaches “A non-transitory computer-readable medium comprising computer executable instructions to cause at least one processor circuit to, for a convolutional layer of the CNN model” (paragraph 8 “Further embodiments relate to apparatuses configured to perform the methods described herein as well as non-transitory computer-readable mediums comprising computer-executable instructions that, when executed by a processor of a device” Lin discloses that further embodiments relate to apparatuses configured to perform the methods using non-transitory computer-readable mediums comprising computer-executable instructions executed by a processor of a device, wherein a person of ordinary skill in the art would have been able to configure the method by Sun/Lin to be performed using such non-transitory computer-readable mediums comprising computer-executable instructions executed by a processor of a device.)
The claim is further rejected under the same rationale as claim 21. The applicant is further directed to the rejection of claim 21 above, because the claim recites similar limitations and processing steps.
Regarding claim 32 depends on claim 31, thus the rejection of claim 31 is incorporated. The claim is further rejected under the same rationale as claim 22. The applicant is further directed to the rejection of claim 22 above, because the claim recites similar limitations and processing steps.
Regarding claim 33 depends on claim 32, thus the rejection of claim 32 is incorporated. The claim is further rejected under the same rationale as claim 23. The applicant is further directed to the rejection of claim 23 above, because the claim recites similar limitations and processing steps.
Regarding claim 34 depends on claim 31, thus the rejection of claim 31 is incorporated. The claim is further rejected under the same rationale as claim 24. The applicant is further directed to the rejection of claim 24 above, because the claim recites similar limitations and processing steps.
Regarding claim 35 depends on claim 31, thus the rejection of claim 31 is incorporated. The claim is further rejected under the same rationale as claim 25. The applicant is further directed to the rejection of claim 25 above, because the claim recites similar limitations and processing steps.
Regarding claim 36 depends on claim 35, thus the rejection of claim 35 is incorporated. The claim is further rejected under the same rationale as claim 26. The applicant is further directed to the rejection of claim 26 above, because the claim recites similar limitations and processing steps.
Regarding claim 37 depends on claim 35, thus the rejection of claim 3351 is incorporated. The claim is further rejected under the same rationale as claim 27. The applicant is further directed to the rejection of claim 27 above, because the claim recites similar limitations and processing steps.
Regarding claim 38 depends on claim 31, thus the rejection of claim 31 is incorporated. The claim is further rejected under the same rationale as claim 28. The applicant is further directed to the rejection of claim 28 above, because the claim recites similar limitations and processing steps.
Regarding claim 40 depends on claim 31, thus the rejection of claim 31 is incorporated. The claim is further rejected under the same rationale as claim 30. The applicant is further directed to the rejection of claim 30 above, because the claim recites similar limitations and processing steps.
Claims 29, 39 are rejected under 35 U.S.C. 103 as being unpatentable over Sun et.al (NPL: Weights Having Stable Signs Are Important: Finding Primary Subnetworks and Kernels to Compress Binary Weight Networks) in view of Lin et.al (US 20230086378 A1), further in view of Fleishman et.al (US 20190311248 A1).
Regarding claim 29 depends on claim 21, thus the rejection of claim 21 is incorporated.
Sun/Lin does not teach the limitation “The apparatus of claim 21, wherein the 1-bit convolutional kernel candidates of the 1-bit convolutional kernel subset are randomly selected from a 1-bit convolutional kernel set including 1-bit convolutional kernel candidates”. However, Fleishman teaches or at least suggest this limitation (paragraph 25 “a random filter generator 208 generates one or more random filters 210 based on the parameters. These filters 210 each separately become part of the CNN configuration ... Each filter may be used in different parts of the CNN, whether per channel 214, per layer 216, or for each convergence operation 218. In some embodiments, the method 200 uses different random filters in a single convolution layer, with each layer learning K filters for integer, K, and each filter being randomly sampled”, and paragraph 33 “The term convolution refers to the filtering process that happens at the convolution layer. The convolution layer takes a filter (also called a kernel)” Fleishman discloses generating random filters for a CNN, including using different random filters in a convolutional layer with each filter being randomly sampled, and further explains that a filter used by a convolutional layer is also called a kernel. Thus, Fleishman teaches randomly sampling convolutional kernels for use in a CNN. Sun discloses a plurality of possible binary kernels and selecting a limited number of those kernels for convolutional-layer quantization. A person of ordinary skill in the art would have understood that applying Fleishman’s random-sampling technique to Sun’s binary-kernel selection would result in randomly selecting the binary-kernel candidates from the available possible binary kernels, thereby teaching or at least suggesting the claimed limitation.)
Before the effective filing date, it would have been obvious to a person ordinary skill in the past to combine teaching of a binary kernel quantization method that uses a selected set of binary kernels to quantize convolutional-layer kernels and compress binary-weight networks by Sun, the teaching of refining the weights of each kernel in a convolutional neural network and implementing the convolutional neural network system and method using memory and computer processor with executable program instruction by Lin, with the teaching of method for random sampled convolutions by Fleishman. The motivation to do so is referred to in Fleishman’s disclosure (paragraph 54 “In some embodiments, the random sampled convolutions method 200 utilizes built-in, general purpose accelerators' capability that provides improved results over other CNNs”, and paragraph 60 “There are some disadvantages to these approaches to save computation cost. Dilated convolutions simply extend the sampling grid, but do so with a fixed factor for all pixels and hence is not as flexible as the random sampled convolutions method … the random sampled convolutions method 200 seamlessly fit the existing training procedure and, in some embodiments, make the convergence happen sooner … the random sampled convolutions method 200, as described above, allows the user to define a non-conventional number of input pixels.” Fleishman discloses that random-sampled convolutions can be incorporated into existing CNN training procedures, provide greater flexibility than fixed sampling approaches, and, in some embodiments, improve CNN performance and accelerate convergence. Accordingly, a person of ordinary skill in the art would have been motivated to apply Fleishman’s random-sampling technique to Sun’s binary-kernel quantization method as an alternative technique for selecting Sun’s binary kernels, while retaining Sun’s binary-kernel quantization framework and obtaining the flexibility and training benefits taught by Fleishman.)
Regarding claim 39 depends on claim 31, thus the rejection of claim 31 is incorporated. The claim is further rejected under the same rationale as claim 29. The applicant is further directed to the rejection of claim 2930 above, because the claim recites similar limitations and processing steps.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DUY TU DIEP whose telephone number is (703)756-1738. The examiner can normally be reached M-F 8-4:30.
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, 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.
/DUY T DIEP/ Examiner, Art Unit 2123
/ALEXEY SHMATOV/ Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights