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 § 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 3-4 and 13-14 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.
Claim 3 and analogous claim 13 recites the limitation "update the first layer by multiplying channel-wise scaling data." There is insufficient antecedent basis for this limitation in the claims.
Claim 3 and analogous claim 13 further recites “update a second layer by dividing the second layer that is an immediately succeeding layer of the first layer by the channel-wise scaling data for each channel.” It is unclear what values of the second layer are being divided by the channel-wise scaling data.
Claim 4 and analogous claim 14 recites the limitations " the channel-wise shifting data… the multiplication result… the channel-wise addition result." There is insufficient antecedent basis for these limitations in the claims.
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.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claim(s) 1-2, 5-9, 11-12, 15-16 are rejected under 35 U.S.C. 102(a)(1) as being anticipated by US Pub no. US20210365779A1 (“Oh et al.”).
In regards to claim 1 and analogous claim 11,
Oh teaches An electronic apparatus comprising: a memory; and at least one processor connected to the memory and configured to control the electronic apparatus; wherein the at least one processor is configured to: obtain a first neural network model comprising at least one layer that may be quantized,
Examiner’s note: The BRI of the limitation merely encompasses “obtain a first neural network model comprising at least one layer” as the recitation of “that may be quantized” does not necessitate that the first neural network must be quantized.
(Oh, “[0015] In accordance with an aspect of the disclosure, an electronic apparatus is provided. The electronic apparatus includes a memory configured to store an artificial intelligence (AI) model including a plurality of layers and a processor [An electronic apparatus comprising: a memory; and at least one processor connected to the memory and configured to control the electronic apparatus; wherein the at least one processor is configured to: obtain a first neural network model ie AI model comprising at least one layer wherein the AI model includes a plurality of layers (at least one layer)],...”)
Oh teaches obtain test data used as an input of the first neural network model,
Examiner’s note: Examiner interprets the test data as a representative data (input data) as input to the first neural network model wherein the first neural network model is a pre-trained NN model in light of the specification of the instant application (“[00172] Referring to FIG. 17, the processor 120 may obtain feature map data from each of at least one layer included in the first neural network model by inputting the test data (representative data) to the first neural network model (pre-trained NN model (fp32).”)
(Oh, [0024], “The control method includes receiving input data and by using an artificial intelligence model, in a neural network computation process relative to the input data [obtain test data used as an input ie input data of the first neural network model]…”)
(Oh, “[0004] The post-training quantization performs IntN Quantization using a pre-trained Float32 model without re-learning and the quantization speed is fast and learning data is not required.”; wherein since learning data (training data) is not required and Oh discloses a post-training process, Examiner interprets the input data to be the ‘test data’ input to the pre-trained model)
Oh teaches obtain feature map data from each of at least one layer included in the first neural network model by inputting the test data to the first neural network model, obtain information about at least one of scaling or shifting to equalize channel-wise data from the feature map data obtained from each of the at least one layer,
PNG
media_image1.png
617
520
media_image1.png
Greyscale
(Oh, [0086], “The processor 120 may compute the input data with a weight value of each of the plurality of first channels included in the first layer [by inputting the test data to the first neural network model; wherein input data is received in the first layer], and may compute an operation result with the weight value of each of the plurality of first channels with the inverse-scaled composite scale parameter based on a shift scaling factor corresponding to each of the plurality of first channels. The processor 120 may compute the feature map data output [obtain feature map data obtained from each of the at least one layer included in the first neural network model; wherein feature map data output is computed from the first layer with weights in a second layer] from the first layer with a weight value of each of a plurality of second channels included in a second layer subsequent to the first layer, and may compute the computation result with the weight value of each of the plurality of second channels with an inverse-scaled composite scale parameter based on a shift scaling factor [obtain information ie shift scaling factor about at least one of scaling or shifting] corresponding to each of the plurality of second channels.”)
(Oh, “[0009] Cross layer equalization (CLE) may perform pre-processing to apply Si (float) scaling, Si −1 (float) rescaling to output of a front convolution layer and input corresponding to a back convolution layer. A scale is continuously adjusted from a first layer to a last layer of the neural network [to equalize channel-wise data from the feature map data obtained from each of the at least one layer; wherein the scale factor is utilized to adjust the outputs of each layer from the first layer to the last layer through cross layer equalization (see also para. [0111-0118] and fig. 1C)], and the whole part is repeated again until there is no change in the Si,...”)
Oh teaches obtain a second neural network model in which channel-wise data of the feature map data is equalized by updating each of the at least one layer based on the obtained information,
Examiner’s note: Examiner interprets the second neural network model as the updated first neural network model in light of the specification ([0086], “Through this operation, the first neural network model may be updated to the second neural network model.”)
(Oh, “[0009] Cross layer equalization (CLE) may perform pre-processing to apply Si (float) scaling, Si −1 (float) rescaling to output of a front convolution layer and input corresponding to a back convolution layer. A scale is continuously adjusted from a first layer to a last layer of the neural network [obtain a second neural network model ie AI model after equalization in which channel-wise data of the feature map data is equalized by updating each of the at least one layer based on the obtained information; see fig. 1C and para. [0111-0118] teaches weight equalization using the WES module], and the whole part is repeated again until there is no change in the Si,...”)
Oh teaches and obtain a quantized third neural network model corresponding to the first neural network model by quantizing the second neural network model based on the test data.
Examiner’s note: Examiner interprets the third neural network model as the quantized second neural network model in light of the specification ([0087], “The processor 120 may obtain quantized third neural network model corresponding to the first neural network model by quantizing the second neural network model based on test data by executing the quantization module 120-6.”)
(Oh, [0009], “…and a general layer-wise quantization is performed when the pre-processing is finished [obtain a quantized third neural network model ie the AI model after equalization and quantization corresponding to the first neural network model by quantizing the second neural network model based on the test data].”)
In regards to claim 2 and analogous claim 12,
Oh teaches The electronic apparatus of claim 1, Oh teaches wherein the at least one processor is further configured to obtain information about at least one of the scaling or the shifting based on a type of a preceding layer and a succeeding layer of the feature map data obtained from each of the at least one layer.
(Oh, “[0009] Cross layer equalization (CLE) may perform pre-processing to apply Si (float) scaling, Si −1 (float) rescaling to output of a front convolution layer and input corresponding to a back convolution layer [obtain information ie the scale factor Si about at least one of the scaling based on a type of a preceding layer ie front conv layer and a succeeding layer ie back conv layer of the feature map data obtained from each of the at least one layer]. A scale is continuously adjusted from a first layer to a last layer of the neural network, and the whole part is repeated again until there is no change in the Si,”)
In regards to claim 5 and analogous claim 15,
Oh teaches The electronic apparatus of claim 1, Oh teaches wherein the at least one processor is further configured to, in a case of obtaining the information about the scaling and the shifting, based on scaling the feature map data of a fifth layer among the at least one layer: apply scaling to the feature map data of the fifth layer,
(Oh, “[0117] The updated parameters (wi) according to the scaling of the 2{circumflex over ( )}Si [apply scaling to the feature map data of the fifth layer] form by channels may be matched with the minimum value, the maximum value range of the entire layer as much as possible, as shown at a lower portion of FIG. 1C and may have a range optimized for layer-wise quantization. The 2{circumflex over ( )}Si scale may be applied to the bias (bi) in the same manner.”)
Oh teaches and apply shifting to the feature map data to which the scaling is applied.
(Oh, “Alternatively, dual shift computation may be performed by shifting the output value by a shift amount, shifting again the output value by the Ch-wise shift scale [apply shifting to the feature map data to which the scaling is applied] and then multiplying by the multiplier.”)
In regards to claim 6,
Oh teaches The electronic apparatus of claim 1, Oh teaches wherein the at least one processor is further configured to obtain information about at least one of the scaling or the shifting based on a channel having a maximum range among the feature map data obtained from each of the at least one layer.
(Oh, “[0112] The WES module may calculate a range of original parameters for each channel, set a range having a maximum value (channel 22 of an upper portion of FIG. 1C) to a reference range, and may obtain a channel-wise shift scaling factor based on the reference range [configured to obtain information ie a channel-wise shift scaling factor about at least one of the scaling or the shifting based on a channel having a maximum range ie a range having a maximum value among the feature map data obtained from each of the at least one layer]”)
In regards to claim 7,
Oh teaches The electronic apparatus of claim 6, Oh teaches wherein the at least one processor is further configured to, in a case of obtaining the information about the scaling and the shifting, for each of the at least one layer: shift a range of a channel having the maximum range, and scale or shift a range of remaining channels based on the shifted range.
(Oh, “[0111] As shown in upper portion of FIG. 1C, the WES module may obtain the minimum value and the maximum values of the original parameter for each channel, and may obtain a channel-wise shift scaling factor (a channel-wise shift scale value) by determining a reference range to have a minimum quantization error using the minimum value and the maximum value for each channel.
[0112] The WES module may calculate a range of original parameters for each channel, set a range having a maximum value [a channel having the maximum range] (channel 22 of an upper portion of FIG. 1C) to a reference range, and may obtain a channel-wise shift scaling factor based on the reference range.
[0114] The WES module may shift scale each channel with respect to the total range [shift a range of a channel having the maximum range… scale or shift a range of remaining channels based on the shifted range], and then sum up the ratio of the changed range values so that the summed value increases, by applying a gradient-descent method, thereby obtaining channel-wise shift scaling factor.”)
In regards to claim 8,
Oh teaches The electronic apparatus of claim 6, Oh teaches wherein the at least one processor is further configured to, in a case of obtaining the information about the scaling, based on scaling the feature map data of a first layer among the at least one layer, obtain information about the scaling so that a range of remaining channels other than a channel having the maximum range, among the feature map data of the first layer, is smaller by a preset value or more than a range of a channel having the maximum range among the feature map data of the first layer.
(Oh, “[0117] The updated parameters (wi) according to the scaling of the 2{circumflex over ( )}Si form by channels may be matched with the minimum value, the maximum value range of the entire layer as much as possible [obtain information about the scaling so that a range of remaining channels other than a channel having the maximum range, among the feature map data of the first layer, is smaller by a preset value or more than a range of a channel having the maximum range among the feature map data of the first layer; ie matching the channels with the minimum and maximum value ranges that forms the scaling factor], as shown at a lower portion of FIG. 1C and may have a range optimized for layer-wise quantization. The 2{circumflex over ( )}Si scale may be applied to the bias (bi) in the same manner.”)
In regards to claim 9,
Oh teaches The electronic apparatus of claim 1, Oh teaches wherein the at least one processor is further configured to obtain the third neural network model by: quantizing at least one layer included in the second neural network model by channels,
(Oh, [0007], “channel-wise quantization [quantizing at least one layer included in the second neural network model by channels] calculates a pair of a minimum value and a maximum value by parameters by proceeding quantization in a channel unit included in a layer. For example, n [min, max] may be obtained for n channels.”)
Oh teaches and quantizing feature map data of each of the at least one layer included in the second neural network model by feature map data.
(Oh, “[0092] Here, (qin, sin, zin) is feature map data (or input data), a scale parameter of a previous layer, and a zero point parameter of a previous layer, as the output data of the previous layers, (qw, sw, zw) is quantized weight value, scale parameter of the weight value, zero point parameter of the weight value, and (qout, sout, zout) is output data of the current layer [quantizing feature map data of each of the at least one layer included in the second neural network model by feature map data], scale parameter of the current layer, and the zero point parameter of the current layer. qi B may represent a quantum bias value after symmetric quantization of floating bias Bi to SintSw scale, and I and j are output channel index and input channel index, respectively.”)
In regards to claim 16,
Oh teaches The method of claim 11, Oh teaches wherein the obtaining of the information about at least one of scaling or shifting comprises detecting an equalization pattern from the obtained feature map data.
Examiner’s note: Examiner interprets the limitation in light of the specification wherein the equalization pattern can be determining whether scaling/shifting is to be applied (“[00177] For example, the processor 120 may identify the equalization pattern of feature map data output from the first layer as scaling if the first layer among the plurality of layers included in the first neural network model is a Conv layer and the second layer immediately after the first layer is the Conv layer”)
(Oh, “[0009] Cross layer equalization (CLE) may perform pre-processing to apply Si (float) scaling, Si −1 (float) rescaling to output of a front convolution layer and input corresponding to a back convolution layer [wherein the obtaining of the information about at least one of scaling comprises detecting an equalization pattern ie determining to apply scaling to a front conv layer and a back conv layer from the obtained feature map data]. A scale is continuously adjusted from a first layer to a last layer of the neural network, and the whole part is repeated again until there is no change in the Si,…”)
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.
Claim(s) 3-4 and 13-14 are rejected under 35 U.S.C. 103 as being unpatentable over Oh in view of Nagel, Markus, et al. "Data-Free Quantization Through Weight Equalization and Bias Correction." arXiv preprint arXiv:1906.04721 (2019). (“Nagel”)
In regards to claim 3 and analogous claim 13,
Oh teaches The electronic apparatus of claim 1,
However, Oh does not explicitly teach wherein the at least one processor is further configured to: in a case of obtaining the information about the scaling, based on scaling the feature map data of a first layer among the at least one layer: update the first layer by multiplying channel-wise scaling data by a weight and a bias of the first layer for each channel, and update a second layer by dividing the second layer that is an immediately succeeding layer of the first layer by the channel-wise scaling data for each channel.
Nagel teaches wherein the at least one processor is further configured to: in a case of obtaining the information about the scaling, based on scaling the feature map data of a first layer among the at least one layer: update the first layer by multiplying channel-wise scaling data by a weight and a bias of the first layer for each channel, and update a second layer by dividing the second layer that is an immediately succeeding layer of the first layer by the channel-wise scaling data for each channel.
(Nagel, Section 4.1.1, fig. 5, “The positive scaling equivariance can be exploited in consecutive layers in neural networks. Given two layers, h = f(W(1)x+b(1)) and y = f(W(2)h+b(2)), through scaling equivariance we have that:
PNG
media_image2.png
103
439
media_image2.png
Greyscale
where S = diag(s) is a diagonal matrix with value Sii denoting the scaling factor si for neuron I [update the first layer by multiplying channel-wise scaling data by a weight and a bias of the first layer; see W(1) and b(1) as the weights and bias of the first layer wherein S are the scaling factors]. This allows us to reparameterize our model with
PNG
media_image3.png
27
147
media_image3.png
Greyscale
PNG
media_image4.png
29
318
media_image4.png
Greyscale
In case of CNNs the scaling will be per channel [for each channel] and broadcast accordingly over the spatial dimensions. The rescaling procedure is illustrated in Figure 5.
PNG
media_image5.png
172
442
media_image5.png
Greyscale
Figure 5. Illustration of the rescaling for a single channel. If scaling factor si scales ci in layer 1; we can instead factor it out and multiply di in layer 2 [update a second layer by dividing the second layer that is an immediately succeeding layer of the first layer by the channel-wise scaling data for each channel; wherein layer 2 (succeeding layer) is updated with an input of ci/si from layer 1 (preceding layer)].”)
Oh and Nagel are both considered to be analogous to the claimed invention because they are in the same field of neural network quantization. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Oh to incorporate the teachings of Nagel in order to provide a data-free quantization method that does not require fine-tuning or hyper-parameter selection, equalizes weights, and improves quantization accuracy performance (Nagel, Abstract, “We introduce a data-free quantization method for deep neural networks that does not require fine-tuning or hyper parameter selection. It achieves near-original model performance on common computer vision architectures and tasks. 8-bit fixed-point quantization is essential for efficient inference on modern deep learning hardware. How ever, quantizing models to run in 8-bit is a non-trivial task, frequently leading to either significant performance reduction or engineering time spent on training a network to be amenable to quantization. Our approach relies on equalizing the weight ranges in the network by making use of a scale-equivariance property of activation functions. In addition the method corrects biases in the error that are introduced during quantization. This improves quantization accuracy performance, and can be applied to many common computer vision architectures with a straight forward API call. For common architectures, such as the MobileNet family, we achieve state-of-the-art quantized model performance. We further show that the method also extends to other computer vision architectures and tasks such as semantic segmentation and object detection.”)
In regards to claim 4 and analogous claim 14,
Oh teaches The electronic apparatus of claim 1,
However, Oh does not explicitly teach wherein the at least one processor is further configured to: in a case of obtaining the information about the shifting, based on shifting the feature map data of a third layer among the at least one layer: update the third layer by adding the channel-wise shifting data to a bias of the third layer by channels, and update a fourth layer by multiplying the channel-wise shifting data by a weight of a fourth layer immediately succeeding the third layer by channels, adding the multiplication result by channels, and subtracting the channel-wise addition result from a bias of the fourth layer.Nagel teaches wherein the at least one processor is further configured to: in a case of obtaining the information about the shifting, based on shifting the feature map data of a third layer among the at least one layer: update the third layer by adding the channel-wise shifting data to a bias of the third layer by channels, and update a fourth layer by multiplying the channel-wise shifting data by a weight of a fourth layer immediately succeeding the third layer by channels, adding the multiplication result by channels, and subtracting the channel-wise addition result from a bias of the fourth layer.
(Nagel, Section 4.1.3, “In case si < 1 the equalization procedure increases bias b(1)i . This could in turn increase the range of the activation quantization. In order to avoid big differences between per channel ranges in the activations we introduce a procedure that absorbs high biases into the subsequent layer.”)
(Nagel, Section 4.1.3, “For a layer with ReLU function r, there is a non-negative vector c such that r(Wx+b-c) = r(Wx+b)-c. The trivial solution c = 0 holds for all x. However, depending on the distribution of x and the values of W and b, there can be some values ci > 0 for which this equality holds for (almost) all x. Following the previous two layer example, these ci can be absorbed from layer 1 into layer 2 as: [update the third layer by adding the channel-wise shifting data ie c to a bias ie b(1) of the third layer by channels;]
PNG
media_image6.png
135
431
media_image6.png
Greyscale
Where
PNG
media_image7.png
27
405
media_image7.png
Greyscale
[update a fourth layer by multiplying the channel-wise shifting data ie c by a weight of a fourth layer immediately succeeding the third layer by channels ie W(2), adding the multiplication result by channels, and subtracting the channel-wise addition result ie c from a bias of the fourth layer ie b(2); wherein Nagel discloses shifting bias as an equalization procedure]
To find c without violating our data-free assumption we assume that the pre-bias activations are distributed normally with the batch normalization shift and scale parameters
PNG
media_image8.png
24
79
media_image8.png
Greyscale
as its mean and standard deviation. We set
PNG
media_image9.png
24
184
media_image9.png
Greyscale
”)
Oh and Nagel are both considered to be analogous to the claimed invention because they are in the same field of neural network quantization. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Oh to incorporate the teachings of Nagel in order to provide a bias equalization procedure as doing so avoids big differences between per channel ranges (Nagel, Section 4.1.3, “In case si < 1 the equalization procedure increases bias b(1)i . This could in turn increase the range of the activation quantization. In order to avoid big differences between per channel ranges in the activations we introduce a procedure that absorbs high biases into the subsequent layer.”)
Claim(s) 10 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Oh in view of Siddegowda, Sangeetha, et al. "Neural network quantization with ai model efficiency toolkit (aimet)." arXiv preprint arXiv:2201.08442 (2022). (“Siddegowda”)
In regards to claim 10,
Oh teaches The electronic apparatus of claim 1,
However, Oh does not explicitly teach wherein the at least one processor is further configured to quantize the second neural network model through affine transformation.Siddegowda teaches wherein the at least one processor is further configured to quantize the second neural network model through affine transformation.
(Siddegowda, Section 2.2, “Uniform affine quantization Uniform quantization is the most commonly used quantization scheme because it permits efficient implementation of fixed-point arithmetic. Uniform affine quantization [quantize the second neural network model through affine transformation], also known as asymmetric quantization, is defined by three quantization parameters: the scale factor s, the zero-point z and the bit-width b. These quantization parameters are also sometimes referred to as quantization encodings. A full set of quantization parameters defines a quantizer. The scale factor and the zero-point are used to to map a floating point value to the integer grid, whose size depends on the bit-width. The scale factor is commonly represented as a floating-point number and specifies the quantization step-size.”)
Oh and Siddegowda are both considered to be analogous to the claimed invention because they are in the same field of neural network quantization. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Oh to incorporate the teachings of Siddegowda in order to provide affine quantization as doing so ensures that real zero is quantized without error and that common operations like zero padding or ReLU do not induce quantization error (Siddegowda, Section 2.2, “…The zero-point is an integer that ensures that real zero is quantized without error. This is important to ensure that common operations like zero padding or ReLU do not induce quantization error.”)
In regards to claim 17,
Oh teaches The method of claim 16, Oh teaches wherein the at least one layer comprises a plurality of layers,
(Oh, “[0015] In accordance with an aspect of the disclosure, an electronic apparatus is provided. The electronic apparatus includes a memory configured to store an artificial intelligence (AI) model including a plurality of layers … [wherein the at least one layer comprises a plurality of layers]”)
However, Oh does not explicitly teach and wherein the detecting of the equalization pattern from the obtained feature map data comprises: comparing each of the plurality of layers with an adjacent layer among the plurality of layers; and determining the equalization pattern based on whether the compared layers are of a convolution layer type, a deconvolution layer type, or a combination of the convolution layer type and the deconvolution layer type.
Siddegowda teaches and wherein the detecting of the equalization pattern from the obtained feature map data comprises: comparing each of the plurality of layers with an adjacent layer among the plurality of layers; and determining the equalization pattern based on whether the compared layers are of a convolution layer type, a deconvolution layer type, or a combination of the convolution layer type and the deconvolution layer type.
(Siddegowda, Section 4.3, “Before applying CLE, the user can use AIMET to visualize the weight ranges of the model and thus determine if a model is a candidate for applying the CLE [comparing each of the plurality of layers with an adjacent layer among the plurality of layers; wherein the user can visualize the weight ranges to determine candidates for CLE wherein determining candidates is interpreted to be comparing the layers (see also Oh, para. [0009] for considering CLE between two conv layers)] technique. For example, if the dynamic range of weights varies significantly across channels of a given convolutional layer, then CLE can be beneficial [determining the equalization pattern based on whether the compared layers are of a convolution layer type; wherein CLE is considered beneficial (ie scaling should be performed) if the layers are conv layers]…”)
Oh and Siddegowda are both considered to be analogous to the claimed invention because they are in the same field of neural network quantization. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Oh to incorporate the teachings of Siddegowda in order to provide a method to visualize weight ranges and determine candidates for applying CLE to allow users to determine whether to apply CLE as doing so would help users understand the effect or improvements after applying CLE to a given model (Siddegowda, Section 4.3, “…Further, visualizations can be used to understand the effect or improvements after applying CLE to a given model.”)
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
NPL: Oh, Jihun, et al. "Weight equalizing shift scaler-coupled post-training quantization." arXiv preprint arXiv:2008.05767 (2020).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JASMINE THAI whose telephone number is (703)756-5904. The examiner can normally be reached M-F 8-4.
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, Michael Huntley can be reached at (303) 297-4307. 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.T.T./Examiner, Art Unit 2129
/MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129
Prosecution Insights