DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 102
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claim(s) 1-4, 8-15, and 17-20 is/are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Lu et al., (US PGPUB 2019/0251441 A1, hereinafter Lu).
Regarding claims 1, 12, and 20, taking claim 20 as exemplary:
Lu shows “A system comprising: a bus; a first processing unit to send a command via the bus; a second processing unit, in response to the command, to: at a layer of a neural network, identify a first channel as a stuck channel based on the first channel having a constant output; and in response to identifying the first channel as a stuck channel, adjust a first operator of the layer.” (Paragraph [0017]: “More specifically, in the various embodiments, a cost metric is determined for a particular layer in the neural network. The cost metric indicates a computational resource cost per channel for the channels included in the particular layer. Because the cost metric indicates a computational resource cost per channel for the channel, the cost metric may be a computation value, or a resource value (or simply a value), for the channel. The cost metric of a layer may be a computation metric of a computational cost of the layer. The neural network is trained via labeled training data. Training the network includes iteratively updating a channel-scaling coefficient for each channel included in the particular layer. A channel-scaling coefficient for a particular channel linearly scales an output of the particular channel. Updating a channel-scaling coefficient for the particular channel is based on the cost metric for the particular layer, as well as other factors discussed herein. Based on the updated channel-scaling coefficients for the channels, channels that provide constant output values independent of their input values (i.e., constant channels) are identified. The trained neural network is then updated to remove (or prune) the constant channels from the particular layer. As such, the updated neural network is a channel-pruned neural network. Although the above discussion contemplates pruning constant channels from a particular layer of the neural network, as discussed herein, channels in multiple layers of the neural network may be similarly pruned.” And in paragraph [0065]: “With reference to FIG. 7, computing device 700 includes a bus 710 that directly or indirectly couples the following devices: memory 712, one or more processors 714, one or more presentation components 716, input/output ports 718, input/output components 720, and an illustrative power supply 722. Bus 710 represents what may be one or more busses (such as an address bus, data bus, or combination thereof).” And in paragraph [0066]: “Computing device 700 typically includes a variety of computer-readable media. Computer-readable media can be any available media that can be accessed by computing device 700 and includes both volatile and nonvolatile media, removable and non-removable media. By way of example, and not limitation, computer-readable media may comprise computer storage media and communication media.” – The constant channel of Lu is the stuck channel and the pruning of the channel of Lu is the adjusting of the operation of the layer.)
Regarding claims 2 and 13, taking claim 2 as exemplary:
Lu shows the method and non-transitory computer readable medium of claims 1 and 12 as claimed and specified above.
And Lu shows “wherein adjusting the first operator comprises: eliminating at least one operation at the first operator.” (Paragraph [0017]: “More specifically, in the various embodiments, a cost metric is determined for a particular layer in the neural network. The cost metric indicates a computational resource cost per channel for the channels included in the particular layer. Because the cost metric indicates a computational resource cost per channel for the channel, the cost metric may be a computation value, or a resource value (or simply a value), for the channel. The cost metric of a layer may be a computation metric of a computational cost of the layer. The neural network is trained via labeled training data. Training the network includes iteratively updating a channel-scaling coefficient for each channel included in the particular layer. A channel-scaling coefficient for a particular channel linearly scales an output of the particular channel. Updating a channel-scaling coefficient for the particular channel is based on the cost metric for the particular layer, as well as other factors discussed herein. Based on the updated channel-scaling coefficients for the channels, channels that provide constant output values independent of their input values (i.e., constant channels) are identified. The trained neural network is then updated to remove (or prune) the constant channels from the particular layer. As such, the updated neural network is a channel-pruned neural network. Although the above discussion contemplates pruning constant channels from a particular layer of the neural network, as discussed herein, channels in multiple layers of the neural network may be similarly pruned.”)
Regarding claims 3 and 14, taking claim 3 as exemplary:
Lu shows the method and non-transitory computer readable medium of claims 2 and 13 as claimed and specified above.
And Lu shows “wherein the at least one operation comprises a multiply operation.” (Paragraph [0033]: “More specifically, cost metric component 222 may determine the cost metric for each layer to be channel-pruned. In various embodiments, one or more of the convolution layers in CNN 240 may be selected for channel pruning. The cost metric for a particular layer indicates a computational resource (such as but not limited to memory) cost per channel for the channels included in particular layer. Because the cost metric indicates a computational resource cost per channel for the channel, the cost metric may be a computation value, or a resource value (or simply a value), for the channel. The cost metric of a layer may be a computation metric of a computational cost of the layer. The cost metric for each convolution layer may be determined via various expressions that indicate a ratio of a measure of the computational cost associated with a particular layer to the number of channels in the layer, prior to pruning of the channels. On exemplary, but non-limiting embodiment is as follows. For the l-th layer, the cost metric (λ.sup.l) may be determined as follows, where l serves as a layer index for CNN 240:.. Furthermore, c.sup.l−1 indicates the channel size (i.e., the number of channels in a layer) of the previous (or upstream-adjacent) layer, over with the convolutions kernel for the l-th convolution layer operates over. Similarly, c.sup.l′ denotes the channel size of follow-up, subsequent, or downstream layers. l.sub.w.sup.l.Math.l.sub.h.sup.l represents the image size of the feature map of the l-th layer.” – The convolution kernels for layer channels are the multiply operations.)
Regarding claims 4 and 15, taking claim 3 as exemplary:
Lu shows the method and non-transitory computer readable medium of claims 2 and 13 as claimed and specified above.
And Lu shows “further comprising: in response to identifying the first channel as a stuck channel, adjusting a second operator of the layer.” (Paragraph [0017]: “More specifically, in the various embodiments, a cost metric is determined for a particular layer in the neural network. The cost metric indicates a computational resource cost per channel for the channels included in the particular layer. Because the cost metric indicates a computational resource cost per channel for the channel, the cost metric may be a computation value, or a resource value (or simply a value), for the channel. The cost metric of a layer may be a computation metric of a computational cost of the layer. The neural network is trained via labeled training data. Training the network includes iteratively updating a channel-scaling coefficient for each channel included in the particular layer. A channel-scaling coefficient for a particular channel linearly scales an output of the particular channel. Updating a channel-scaling coefficient for the particular channel is based on the cost metric for the particular layer, as well as other factors discussed herein. Based on the updated channel-scaling coefficients for the channels, channels that provide constant output values independent of their input values (i.e., constant channels) are identified. The trained neural network is then updated to remove (or prune) the constant channels from the particular layer. As such, the updated neural network is a channel-pruned neural network. Although the above discussion contemplates pruning constant channels from a particular layer of the neural network, as discussed herein, channels in multiple layers of the neural network may be similarly pruned.” And in paragraph [0033]: “More specifically, cost metric component 222 may determine the cost metric for each layer to be channel-pruned. In various embodiments, one or more of the convolution layers in CNN 240 may be selected for channel pruning. The cost metric for a particular layer indicates a computational resource (such as but not limited to memory) cost per channel for the channels included in particular layer. Because the cost metric indicates a computational resource cost per channel for the channel, the cost metric may be a computation value, or a resource value (or simply a value), for the channel. The cost metric of a layer may be a computation metric of a computational cost of the layer. The cost metric for each convolution layer may be determined via various expressions that indicate a ratio of a measure of the computational cost associated with a particular layer to the number of channels in the layer, prior to pruning of the channels. On exemplary, but non-limiting embodiment is as follows. For the l-th layer, the cost metric (λ.sup.l) may be determined as follows, where l serves as a layer index for CNN 240:.. Furthermore, c.sup.l−1 indicates the channel size (i.e., the number of channels in a layer) of the previous (or upstream-adjacent) layer, over with the convolutions kernel for the l-th convolution layer operates over. Similarly, c.sup.l′ denotes the channel size of follow-up, subsequent, or downstream layers. l.sub.w.sup.l.Math.l.sub.h.sup.l represents the image size of the feature map of the l-th layer.”)
Regarding claims 8 and 17, taking claim 8 as exemplary:
Lu shows the method and non-transitory computer readable medium of claims 1 and 12 as claimed and specified above.
And Lu shows “further comprising: in response to identifying the first channel as a stuck channel, adding a second operator to the layer of the neural network.” (Paragraph [0007]: “When a convolution layer is not batch normalized, the model weights of the convolution layer may be transformed, such that the non-batch normalized convolution layer may be channel-pruned. For such convolution layers, the model bias coefficient for the channels may be removed and/or transformed. A scaling coefficient for each channel within the convolution layer may be determined based on a variance of a convolution of mini batches of training data. A batch normalization bias coefficient may be determined for each channel based on a mean of the convolution of the mini batches of the training data.” And in paragraph [0020]: “The embodiments herein include an end-to-end training platform for training a DNN, wherein the training includes iteratively updating the channel-scaling coefficients for channels within at least a portion of the layers of the DNN. The training includes a bias toward closing information gates when the performance of the DNN does not significantly suffer. That is, the embodiments determine the channel-scaling coefficients, wherein a norm of the channel-scaling coefficients is penalized in a BN training loss function. The training of a DNN includes minimizing, or at least decreasing, the BN training loss function, as described herein. More particularly, when the value of a channel-scaling coefficient is zeroed, the output of the channel is a constant based on a BN bias coefficient of the channel. A constant signal is a high-entropy signal that carries no, or at least an insignificant amount of, information. Because the output of the channel is constant, the channel does not contribute to information flowing though the DNN. Thus, a constant channel may be pruned (or removed) from the DNN, and the BN bias coefficient of the constant channel may be absorbed into a subsequent layer, without a significant impact on the performance of the DNN.” – the absorption of a constant channel is adding a second operation to the layer of the network.)
Regarding claims 9 and 18, taking claim 9 as exemplary:
Lu shows the method and non-transitory computer readable medium of claims 8 and 17 as claimed and specified above.
And Lu shows “wherein adding the second operator comprises: adding the second operator to add a bias, the bias being equivalent to the constant output..” (Paragraph [0007]: “When a convolution layer is not batch normalized, the model weights of the convolution layer may be transformed, such that the non-batch normalized convolution layer may be channel-pruned. For such convolution layers, the model bias coefficient for the channels may be removed and/or transformed. A scaling coefficient for each channel within the convolution layer may be determined based on a variance of a convolution of mini batches of training data. A batch normalization bias coefficient may be determined for each channel based on a mean of the convolution of the mini batches of the training data.” And in paragraph [0020]: “The embodiments herein include an end-to-end training platform for training a DNN, wherein the training includes iteratively updating the channel-scaling coefficients for channels within at least a portion of the layers of the DNN. The training includes a bias toward closing information gates when the performance of the DNN does not significantly suffer. That is, the embodiments determine the channel-scaling coefficients, wherein a norm of the channel-scaling coefficients is penalized in a BN training loss function. The training of a DNN includes minimizing, or at least decreasing, the BN training loss function, as described herein. More particularly, when the value of a channel-scaling coefficient is zeroed, the output of the channel is a constant based on a BN bias coefficient of the channel. A constant signal is a high-entropy signal that carries no, or at least an insignificant amount of, information. Because the output of the channel is constant, the channel does not contribute to information flowing though the DNN. Thus, a constant channel may be pruned (or removed) from the DNN, and the BN bias coefficient of the constant channel may be absorbed into a subsequent layer, without a significant impact on the performance of the DNN.” And in paragraph [0017]: “A channel-scaling coefficient for a particular channel linearly scales an output of the particular channel. Updating a channel-scaling coefficient for the particular channel is based on the cost metric for the particular layer, as well as other factors discussed herein. Based on the updated channel-scaling coefficients for the channels, channels that provide constant output values independent of their input values (i.e., constant channels) are identified. The trained neural network is then updated to remove (or prune) the constant channels from the particular layer.”.)
Regarding claims 10 and 19, taking claim 10 as exemplary:
Lu shows the method and non-transitory computer readable medium of claims 1 and 12 as claimed and specified above.
And Lu shows “further comprising: identifying the first channel has the constant output based on application of a range of input values to the first channel.” (Paragraph [0017]: “More specifically, in the various embodiments, a cost metric is determined for a particular layer in the neural network. The cost metric indicates a computational resource cost per channel for the channels included in the particular layer. Because the cost metric indicates a computational resource cost per channel for the channel, the cost metric may be a computation value, or a resource value (or simply a value), for the channel. The cost metric of a layer may be a computation metric of a computational cost of the layer. The neural network is trained via labeled training data. Training the network includes iteratively updating a channel-scaling coefficient for each channel included in the particular layer. A channel-scaling coefficient for a particular channel linearly scales an output of the particular channel. Updating a channel-scaling coefficient for the particular channel is based on the cost metric for the particular layer, as well as other factors discussed herein. Based on the updated channel-scaling coefficients for the channels, channels that provide constant output values independent of their input values (i.e., constant channels) are identified. The trained neural network is then updated to remove (or prune) the constant channels from the particular layer. As such, the updated neural network is a channel-pruned neural network. Although the above discussion contemplates pruning constant channels from a particular layer of the neural network, as discussed herein, channels in multiple layers of the neural network may be similarly pruned.”)
Regarding claim 11:
Lu shows the method of claim 11 as claimed and specified above.
And Lu shows “wherein identifying the first channel has the constant output comprises: applying the range of input values to a first operator of the layer to generate a first range of output values; applying a minimum and maximum of the first range of output values as inputs to a second operator to generate a second range of output values; and identifying the first channels has the constant output based on the second range of output values.” (Paragraph [0017]: “More specifically, in the various embodiments, a cost metric is determined for a particular layer in the neural network. The cost metric indicates a computational resource cost per channel for the channels included in the particular layer. Because the cost metric indicates a computational resource cost per channel for the channel, the cost metric may be a computation value, or a resource value (or simply a value), for the channel. The cost metric of a layer may be a computation metric of a computational cost of the layer. The neural network is trained via labeled training data. Training the network includes iteratively updating a channel-scaling coefficient for each channel included in the particular layer. A channel-scaling coefficient for a particular channel linearly scales an output of the particular channel. Updating a channel-scaling coefficient for the particular channel is based on the cost metric for the particular layer, as well as other factors discussed herein. Based on the updated channel-scaling coefficients for the channels, channels that provide constant output values independent of their input values (i.e., constant channels) are identified. The trained neural network is then updated to remove (or prune) the constant channels from the particular layer. As such, the updated neural network is a channel-pruned neural network. Although the above discussion contemplates pruning constant channels from a particular layer of the neural network, as discussed herein, channels in multiple layers of the neural network may be similarly pruned.” And in paragraph [0042]: “Channel pruning component 232 identifies constant channels in the layers based on the trained values of the channel-scaling coefficients. That is, channels where the channel-scaling coefficients (γ.sub.k.sup.l) have converged to 0.0 via training by BN scaling coefficients trainer 230 are identified. Channel pruning component 232 prunes or removes those identified constant channels from the CNN. Channel pruning component 232 also updates subsequent layers in the CNN, such that the updated subsequent layers absorb the constant channels that have been pruned from the previous layers. More particularly, a follow-up or adjacent-downstream layer (l+1) of a channel-pruned layer (l) is updated to absorb the BN bias coefficient of a pruned channel in the channel-pruned layer. Absorbing the constant channels into subsequent layers is based on whether the subsequent layers is a batch normalized layer. As noted throughout, for the functionality of the CNN is to not be negatively affected, the BN bias coefficients or a pruned channel are absorbed into the (batch normalized or not batch normalized) subsequent layer. For the l-th BN convolution layer that is subject to a rectified linear unit (ReLU) and channel pruning, the output (x.sup.l+1) based on the input (x.sup.l)” in paragraph [0053]: “At block 308, constant channels included in the convolution layers are identified based on the updated channel-scaling coefficients. For example, channel pruning component 232 of FIG. 2 may identify channels with constant output, as indicated by a zero-valued channel-scaling coefficient. At block 310, the identified constant channels are pruned (or removed) from the NN. In various embodiments, the channel pruning component 232 may update the trained NN by removing the constant channels form the layers, such that the updated NN is a channel-pruned NN. At block 312, the channel-pruned NN is provided. For example, the channel-pruned NN may be provided to NN computing device 104 of FIG. 1.” And in paragraph [0055]: “At block 404, the NN is iteratively trained based on the cost metrics (as determined via block 302 of FIG. 3), as well as the scaled BN scaling coefficients and model weights (as scaled via block 402). Various embodiments of iteratively training the NN are discussed in conjunction with network trainer component 226 of FIG. 2 and process 500 of FIG. 5A. However, briefly here, the model weights of the NN may be trained via stochastic gradient descent (SGD) of a training loss function. The BN scaling coefficients are updated via an iterative-thresholding algorithm (ISTA) that penalizes a batch normalization loss function based on the cost metrics and a norm of the BN scaling coefficients. At block 406, constant channels of the convolution layers are identified and removed based on the trained BN scaling coefficients of the channels. Various embodiments of identifying and removing constant channels are discussed in conjunction with at channel-pruning component 232 of FIG. 2 and process 500 of FIG. 5A.”)
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claim(s) 5-7 and 16 is/are rejected under 35 U.S.C. 103 as being unpatentable over Lu in view of Trusov et al., (US 2024/0211763 A1, hereinafter Trusov).
Regarding claims 5 and 16, taking claim 5 as exemplary:
Lu teaches the method and non-transitory computer readable of claims 4 and 15 as claimed and specified above.
But Lu does not appear to explicitly recite “wherein the first operator is a quantized or non-quantized operator, and the second operator is a quantized operator.”
However, Trusov teaches “wherein the first operator is a quantized or non-quantized operator, and the second operator is a quantized operator.” (Paragraph [0012]: “The method may further comprise using the at least one hardware processor to quantize each of the one or more layers to be quantized in the neural network by: training the layer without quantizing inputs to the layer; collecting a histogram of the inputs to the layer; determining input quantization parameters that minimize quantization error for the inputs to the layer based on the histogram; quantizing the inputs to the layer channel by channel using the input quantization parameters; determining weight quantization parameters that minimize quantization error for weights of the layer; quantizing the weights of the layer filter by filter using the weight quantization parameters; and quantize a bias of the layer based on one or both of the input quantization parameters and the weight quantization parameters. In an embodiment, the input quantization parameters are frozen after being determined. In an embodiment, the weights and the bias of the layer are not frozen during the quantization. The method may further comprise using the at least one hardware processor to, during quantization of each of the one or more layers to be quantized in the neural network, fine-tune the layer: after quantizing the inputs to the layer and before determining the weight quantization parameters; after quantizing the weights of the layer and before quantizing the bias of the layer; and after quantizing the bias of the layer.”)
Lu and Trusov are analogous in the arts because both Lu and Trusov describe channel layer data operations of neural networks.
Therefore, it would be obvious to one of ordinary skill in the art at the filing date of the instant application, having the teachings of Lu and Trusov before him or her, to modify the teachings of Lu to include the teachings of Trusov in order to increase efficiency of neural network operations on computational devices who “may not be able to provide high speed processing” (see Trusov paragraph [0004]) through quantization (see Trusov paragraphs [0011] and [0012]).
Regarding claim 6:
Lu and Trusov teach the method of claim 5 as claimed and specified above.
And Lu shows “wherein the first operator provides data to an input of the second operator.” (Paragraph [0017]: “More specifically, in the various embodiments, a cost metric is determined for a particular layer in the neural network. The cost metric indicates a computational resource cost per channel for the channels included in the particular layer. Because the cost metric indicates a computational resource cost per channel for the channel, the cost metric may be a computation value, or a resource value (or simply a value), for the channel. The cost metric of a layer may be a computation metric of a computational cost of the layer. The neural network is trained via labeled training data. Training the network includes iteratively updating a channel-scaling coefficient for each channel included in the particular layer. A channel-scaling coefficient for a particular channel linearly scales an output of the particular channel. Updating a channel-scaling coefficient for the particular channel is based on the cost metric for the particular layer, as well as other factors discussed herein. Based on the updated channel-scaling coefficients for the channels, channels that provide constant output values independent of their input values (i.e., constant channels) are identified. The trained neural network is then updated to remove (or prune) the constant channels from the particular layer. As such, the updated neural network is a channel-pruned neural network. Although the above discussion contemplates pruning constant channels from a particular layer of the neural network, as discussed herein, channels in multiple layers of the neural network may be similarly pruned.” And in paragraph [0033]: “More specifically, cost metric component 222 may determine the cost metric for each layer to be channel-pruned. In various embodiments, one or more of the convolution layers in CNN 240 may be selected for channel pruning. The cost metric for a particular layer indicates a computational resource (such as but not limited to memory) cost per channel for the channels included in particular layer. Because the cost metric indicates a computational resource cost per channel for the channel, the cost metric may be a computation value, or a resource value (or simply a value), for the channel. The cost metric of a layer may be a computation metric of a computational cost of the layer. The cost metric for each convolution layer may be determined via various expressions that indicate a ratio of a measure of the computational cost associated with a particular layer to the number of channels in the layer, prior to pruning of the channels. On exemplary, but non-limiting embodiment is as follows. For the l-th layer, the cost metric (λ.sup.l) may be determined as follows, where l serves as a layer index for CNN 240:.. Furthermore, c.sup.l−1 indicates the channel size (i.e., the number of channels in a layer) of the previous (or upstream-adjacent) layer, over with the convolutions kernel for the l-th convolution layer operates over. Similarly, c.sup.l′ denotes the channel size of follow-up, subsequent, or downstream layers. l.sub.w.sup.l.Math.l.sub.h.sup.l represents the image size of the feature map of the l-th layer.”)
Regarding claim 7:
Lu and Trusov teach the method of claim 5 as claimed and specified above.
But Lu does not appear to explicitly recite “wherein the second operator provides data to an input of the first operator.”
However, Trusov teaches “wherein the second operator provides data to an input of the first operator.” (Paragraph [0012]: “The method may further comprise using the at least one hardware processor to quantize each of the one or more layers to be quantized in the neural network by: training the layer without quantizing inputs to the layer; collecting a histogram of the inputs to the layer; determining input quantization parameters that minimize quantization error for the inputs to the layer based on the histogram; quantizing the inputs to the layer channel by channel using the input quantization parameters; determining weight quantization parameters that minimize quantization error for weights of the layer; quantizing the weights of the layer filter by filter using the weight quantization parameters; and quantize a bias of the layer based on one or both of the input quantization parameters and the weight quantization parameters. In an embodiment, the input quantization parameters are frozen after being determined. In an embodiment, the weights and the bias of the layer are not frozen during the quantization. The method may further comprise using the at least one hardware processor to, during quantization of each of the one or more layers to be quantized in the neural network, fine-tune the layer: after quantizing the inputs to the layer and before determining the weight quantization parameters; after quantizing the weights of the layer and before quantizing the bias of the layer; and after quantizing the bias of the layer.”)
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
Nie et al., (US PGPUB 2021/0287092 A1), part of the prior art made of record, teaches the adjusting of channels of layers of a network for changing operations of a channel of claims 1, 12, and 20 in paragraph [0025] through the use of pruning of channels of a layer.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SHANE D WOOLWINE whose telephone number is (571)272-4138. The examiner can normally be reached M-F 9:30-6:00 PM.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, MIRANDA HUANG can be reached at (571) 270-7092. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
SHANE D. WOOLWINE
Primary Examiner
Art Unit 2124
/SHANE D WOOLWINE/Primary Examiner, Art Unit 2124
Prosecution Insights