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 .
Response to Arguments
Applicant's arguments filed 10/20/25 have been fully considered but they are not fully persuasive.
Regarding the 101 rejections, applicant’s arguments and amendments to the independent claims are persuasive and overcome the previous 101 rejections. Specifically, applicant’s amended limitations process the data representing the compressed set of weight values, using the weight decoder selection circuitry, to determine a type of layer of the neural network for which the compressed set of weight values are to be used; select a weight decoder from the set of weight decoders using the weight decoder selection circuitry, wherein selecting a weight decoder is based on the type of layer of the neural network for which the compressed set of weight values are to be used; process the data representing the compressed set of weight values using the selected weight decoder to obtain the uncompressed set of weight values; and provide the set of uncompressed weight values to the convolution engine provides a technical improvement to reducing convolutional neural network’s resource usage by compressing and decompressing weight data for specific layers of the CNN. See pg. 12-13 of “Remarks”: “The claimed system solves this problem through a specific hardware configuration where:
Both decoders can decompress the same compressed weight values. This is not merely selecting between different data formats, but rather a specific technical architecture where a single compressed dataset is deliberately designed to be decodable by multiple hardware decoders using different decompression functions.
The decoders produce outputs in different weight value spaces. The first decoder produces uncompressed weights in a first (larger) weight value space providing higher precision, while the second decoder produces uncompressed weights in a second, smaller weight value space providing lower precision. This is a deliberate hardware design choice that enables runtime flexibility.
Dynamic runtime adaptation is performed based on layer type. The system processes the compressed weight data to determine the type of neural network layer, then selects the appropriate decoder based on that layer type. This enables the system to dynamically adapt to runtime constraints by choosing between:
Higher precision/higher computational cost (first decoder) for layers where few weights are decoded and reused many times
Lower precision/lower computational cost (second decoder) for layers requiring many weight values”
Applicant’s amendments and corresponding arguments that the claimed invention provides a technical improvement to the field of CNN resource usage are persuasive. Therefore, the 101 rejections are withdrawn.
Regarding the 103 rejections, applicant's arguments filed with respect to the prior art rejections have been fully considered but they are moot. Applicant has amended the claims to recite new combinations of limitations. Applicant's arguments are directed at the amendment. Please see below for new grounds of rejection, necessitated by Amendment.
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 1-4, 9, 11, 13-15, and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Chalfin, et al., US Pre-Grant Publication 2018/0239992A1 (“Chalfin”) in view of Ross, et al., US Pre-Grant Publication 2019/0207626A1 (“Ross”) and further in view of Holland, US Pre-Grant Publication 2019/0228284A1 (“Holland”).
Regarding claim 1 and analogous claims 13 and 20, Chalfin discloses:
A data processing system comprising: a convolution engine for implementing one or more layers of a neural network, the convolution engine being configured to receive an uncompressed set of weight values and an input feature map and to perform convolution operations using the uncompressed set of weight values and the input feature map to produce an output feature map; (Chalfin, ⁋83, “the artificial neural network may comprise a convolutional neural network [A data processing system comprising: a convolution engine], for example having one or more convolutional layers [for implementing one or more layers of a neural network,] (e.g. that apply one or more convolution operations), one or more pooling layers (e.g. that pool or aggregate sets of input values to generate an output), and/or one or more fully connected layers (e.g. comprising one or more activation layers). Weight values may be applied at one or more convolutional layers when producing a result from an input to the artificial neural network [and an input feature map and to perform convolution operations using the uncompressed set of weight values and the input feature map to produce an output feature map;].” and Chalfin, ⁋29, “These embodiments can therefore significantly reduce the amount of processing resources needed, for example when it is desired to apply or process only selected decompressed weight values in the artificial neural network [to receive an uncompressed set of weight values].”).
…wherein the data processing system is configured to: obtain data representing the compressed set of weight values; (Chalfin, ⁋22, “use an image compression scheme to compress the array of weight values to provide compressed weight data for the artificial neural network […wherein the data processing system is configured to: obtain data representing the compressed set of weight values;].”).
and provide the set of uncompressed weight values to the convolution engine. (Chalfin, ⁋29, “These embodiments can therefore significantly reduce the amount of processing resources needed, for example when it is desired to apply or process only selected decompressed weight values in the artificial neural network; the neural network is interpreted as a convolutional neural network (i.e. and provide the set of uncompressed weight values to the convolution engine.)”).
While Chalfin teaches compressing model weights of a convolutional model, Chalfin does not explicitly teach:
a set of weight decoders comprising: a first weight decoder configured to decompress a compressed set of weight values using a first decompression function to obtain the uncompressed set of weight values,
wherein weight values of the uncompressed set of weight values obtained using the first decompression function are represented in a first weight value space;
a second weight decoder configured to decompress the compressed set of weight values using a second decompression function to obtain the uncompressed set of weight values,
wherein weight values of the uncompressed set of weight values obtained using the second decompression function are represented in a second, smaller, weight value space;
and a weight decoder selection circuitry for selecting a weight decoder from the set of weight decoders,…
wherein the compressed set of weight values are capable of being decompressed using the first weight decoder and capable of being decompressed using the second weight decoder;
process the data representing the compressed set of weight values, using the weight decoder selection circuitry, to determine a type of layer of the neural network for which the compressed set of weight values are to be used;
select a weight decoder from the set of weight decoders using the weight decoder selection module;
wherein selecting a weight decoder is based on the type of layer of the neural network for which the compressed set of weight values are to be used;
process the data representing the compressed set of weight values using the selected weight decoder to obtain the uncompressed set of weight values;
Ross teaches:
a set of weight decoders comprising: a first weight decoder configured to decompress a compressed set of weight values using a first decompression function to obtain the uncompressed set of weight values, (Ross, ⁋29, “a different decompression circuit (e.g., decompression circuits 110A and 110B) [a set of weight decoders] may be used to decompress the different sets of compressed model coefficients compressed using different functions, to produce different sets of decompressed model coefficients (e.g., decompressed model coefficients 116A and 116B) [a first weight decoder configured to decompress a compressed set of weight values using a first decompression function to obtain the uncompressed set of weight values,].”).
a second weight decoder configured to decompress the compressed set of weight values using a second decompression function to obtain the uncompressed set of weight values, (Ross, ⁋29, “a different decompression circuit (e.g., decompression circuits 110A and 110B) may be used to decompress the different sets of compressed model coefficients compressed using different functions, to produce different sets of decompressed model coefficients (e.g., decompressed model coefficients 116A and 116B) [a second weight decoder configured to decompress the compressed set of weight values using a second decompression function to obtain the uncompressed set of weight values,].”).
and a weight decoder selection circuitry for selecting a weight decoder from the set of weight decoders,… (Ross, ⁋32, “In some embodiments, the compiler, when compressing the model coefficients to be stored in the memory of the TSP, determines which functions are used for compression/decompression of the model coefficients for which bit channels and at which times [and a weight decoder selection module for selecting a weight decoder from the set of weight decoders,…].”).
select a weight decoder from the set of weight decoders using the weight decoder selection circuitry; (Ross, ⁋32, “In some embodiments, the compiler, when compressing the model coefficients to be stored in the memory of the TSP, determines which functions are used for compression/decompression of the model coefficients for which bit channels and at which times [select a weight decoder from the set of weight decoders using the weight decoder selection circuitry;].”).
process the data representing the compressed set of weight values using the selected weight decoder to obtain the uncompressed set of weight values; (Ross, ⁋29, “a different decompression circuit (e.g., decompression circuits 110A and 110B) may be used to decompress the different sets of compressed model coefficients [process the data representing the compressed set of weight values] compressed using different functions, to produce different sets of decompressed model coefficients (e.g., decompressed model coefficients 116A and 116B) [using the selected weight decoder to obtain the uncompressed set of weight values;].”).
Chalfin and Ross are both in the same field of endeavor (i.e. model compression). It would have been obvious for a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chalfin and Ross to teach the above limitation(s). The motivation for doing so is that decompression functions reduce memory usage (cf. Ross, ⁋20, “The use of decompression functions also reduces the amount of memory look-ups required during decompression.”).
While Chalfin in view of Ross teaches a convolutional model that compresses and decompresses model weights, the combination does not explicitly teach:
wherein weight values of the uncompressed set of weight values obtained using the first decompression function are represented in a first weight value space; (Holland, ⁋61, “For each layer of the computational network (e.g., DCN 350), a different compression rate may be specified. For example, a lossless compression may be applied for weights 502 of layer 1 and weights 506 of layer #2 [the first decompression function], while a lossy compression may be applied to weights in layer N. In lossless compression, each unit (e.g., within weights 502), which may be made out of a fixed memory size (e.g., 256 Bytes), is compressed or reduced to a smaller or minimum possible size without loss of features [wherein weight values of the uncompressed set of weight values obtained using the first decompression function are represented in a first weight value space;].”).
wherein weight values of the uncompressed set of weight values obtained using the second decompression function are represented in a second, smaller, weight value space; (Holland, ⁋61, “For each layer of the computational network (e.g., DCN 350), a different compression rate may be specified. For example, a lossless compression may be applied for weights 502 of layer 1 and weights 506 of layer #2, while a lossy compression [the second decompression function] may be applied to weights in layer N…In lossy compression, a unit is compressed (e.g., a unit of activations 504) to a level in which there some loss of features. That is, the units are compressed or reduced to a smaller or minimum possible size such that the units may be later decompressed to be approximately equal, but not identical to the unit prior to compression; lossy compression/decompression is interpreted as being smaller than lossless as lossy trades size reduction for accuracy while lossless does not compromise accuracy (i.e. wherein weight values of the uncompressed set of weight values obtained using the second decompression function are represented in a second, smaller, weight value space;).”).
wherein the compressed set of weight values are capable of being decompressed using the first weight decoder and capable of being decompressed using the second weight decoder; (Holland, ⁋61, “For each layer of the computational network (e.g., DCN 350) [wherein the compressed set of weight values], a different compression rate may be specified. For example, a lossless compression [are capable of being decompressed using the first weight decoder] may be applied for weights 502 of layer 1 and weights 506 of layer #2, while a lossy compression [and capable of being decompressed using the second weight decoder;] may be applied to weights in layer N.”).
process the data representing the compressed set of weight values, using the weight decoder selection circuitry, to determine a type of layer of the neural network for which the compressed set of weight values are to be used; (Holland, ⁋63, “The compression for the weight or activations may be determined based on a compression ratio. The compression ratio (e.g., 50%, no compression) may be specified based on a compression map (e.g., such as a compression map produced via compression map unit 402). The compression map may be generated based on a system event such as a bandwidth condition, a power condition, a debug condition, or a thermal condition, for example. In some aspects, the compression map may be determined based on the type of layer (e.g., convolution layer or pooling layer) [process the data representing the compressed set of weight values, using the weight decoder selection circuitry, to determine a type of layer of the neural network for which the compressed set of weight values are to be used;]”).
wherein selecting a weight decoder is based on the type of layer of the neural network for which the compressed set of weight values are to be used; (Holland, ⁋63, “The compression for the weight or activations may be determined based on a compression ratio. The compression ratio (e.g., 50%, no compression) may be specified based on a compression map (e.g., such as a compression map produced via compression map unit 402). The compression map may be generated based on a system event such as a bandwidth condition, a power condition, a debug condition, or a thermal condition, for example. In some aspects, the compression map may be determined based on the type of layer (e.g., convolution layer or pooling layer) [wherein selecting a weight decoder is based on the type of layer of the neural network for which the compressed set of weight values are to be used;]”).
Chalfin, in view of Ross, and Holland are both in the same field of endeavor (i.e. model compression). It would have been obvious for a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chalfin, in view of Ross, and Holland to teach the above limitation(s). The motivation for doing so is that adapting compression level per layer improves the quality and performance of a neural network (cf. Holland, ⁋10, “To address the issues of quality loss and increased bandwidth and power consumption in a computational system, aspects of the present disclosure are directed to applying a mixture of lossless and lossy compression to neural network layers as the layers are written to or read from a chip memory.”).
Regarding claim 2, Chalfin in view of Ross and Holland teaches the data processing system of claim 1. Ross further teaches wherein the first weight decoder and the second weight decoder are different in at least one respect. (Ross, ⁋29, “In some embodiments, a different decompression circuit (e.g., decompression circuits 110A [wherein the first weight decoder] and 110B [and the second weight decoder]) may be used to decompress the different sets of compressed model coefficients compressed using different functions; using different decompression functions for different compression functions is interpreted as the weight decoders being different (i.e. are different in at least one respect.), to produce different sets of decompressed model coefficients (e.g., decompressed model coefficients 116A and 116B).”).
It would have been obvious to one of ordinary skill in the art before the effective filling date of the present application to combine the teachings of Ross with the teachings of Chalfin and Holland for the same reasons disclosed in claim 1.
Regarding claim 3 and analogous claim 14, Chalfin in view of Ross and Holland teaches the data processing system of claim 2. Ross further teaches wherein the first weight decoder is associated with a first characteristic, the second weight decoder is associated with a second characteristic, and the first characteristic is different to the second characteristic. (Ross, ⁋29, “In some embodiments, a different decompression circuit (e.g., decompression circuits 110A [wherein the first weight decoder] and 110B [and the second weight decoder]) may be used to decompress the different sets of compressed model coefficients compressed using different functions; using different decompression functions for different compression functions is interpreted as the weight decoders being associated with different characteristics (i.e. wherein the first weight decoder is associated with a first characteristic, the second weight decoder is associated with a second characteristic, and the first characteristic is different to the second characteristic.), to produce different sets of decompressed model coefficients (e.g., decompressed model coefficients 116A and 116B).”).
It would have been obvious to one of ordinary skill in the art before the effective filling date of the present application to combine the teachings of Ross with the teachings of Chalfin and Holland for the same reasons disclosed in claim 1.
Regarding claim 4 and analogous claim 15, Chalfin in view of Ross and Holland teaches the data processing system of claim 3. Holland further teaches wherein the first characteristic includes a first compression ratio and the second characteristic includes a second compression ratio, and the first compression ratio is different to the second compression ratio. (Holland, ⁋61, “For each layer of the computational network (e.g., DCN 350), a different compression rate may be specified [and the first compression ratio is different to the second compression ratio.]. For example, a lossless compression [wherein the first characteristic includes a first compression ratio] may be applied for weights 502 of layer 1 and weights 506 of layer #2, while a lossy compression [and the second characteristic includes a second compression ratio,] may be applied to weights in layer N.”).
It would have been obvious to one of ordinary skill in the art before the effective filling date of the present application to combine the teachings of Holland with the teachings of Chalfin and Ross for the same reasons disclosed in claim 1.
Regarding claim 9, Chalfin in view of Ross and Holland teaches the data processing system of claim 1. Holland further teaches wherein the data representing the compressed set of weight values comprises an indication of which of the set of weight decoders is to be selected. (Holland, ⁋61, “For each layer of the computational network (e.g., DCN 350) [wherein the data representing the compressed set of weight values], a different compression rate may be specified. For example, a lossless compression may be applied for weights 502 of layer 1 and weights 506 of layer #2, while a lossy compression may be applied to weights in layer N; the layers determine which compression/decompression is selected (i.e. comprises an indication of which of the set of weight decoders is to be selected.).”).
It would have been obvious to one of ordinary skill in the art before the effective filling date of the present application to combine the teachings of Holland with the teachings of Chalfin and Ross for the same reasons disclosed in claim 1.
Regarding claim 11 and analogous claim 19, Chalfin in view of Ross and Holland teaches the data processing system of claim 1. Ross further teaches wherein at least one of the first decompression function and the second decompression function comprises using a lookup table to decompress the compressed set of weight values. (Ross, ⁋18, “To reduce memory usage, the set of coefficients are compressed prior to storage. The stored compressed coefficients will need to be decompressed prior to operating on the input data. Look-up tables may be used to map compressed coefficient values to decompressed coefficient values [wherein at least one of the first decompression function and the second decompression function comprises using a lookup table to decompress the compressed set of weight values.].”).
It would have been obvious to one of ordinary skill in the art before the effective filling date of the present application to combine the teachings of Ross with the teachings of Chalfin and Holland for the same reasons disclosed in claim 1.
Regarding claim 20, the claim is analogous to claim 1. Chalfin further teaches the additional limitations A non-transitory computer-readable storage medium comprising computer- executable instructions which, when executed by one or more processor, cause the processors to: (Chalfin, claim 21, “A non-transitory computer readable storage medium storing computer software code which when executing on a processor [A non-transitory computer-readable storage medium comprising computer- executable instructions which, when executed by one or more processor, cause the processors to:]”).
Claims 7 and 10 are rejected under 35 U.S.C. 103 as being unpatentable over Chalfin, et al., US Pre-Grant Publication 2018/0239992A1 (“Chalfin”) in view of Ross, et al., US Pre-Grant Publication 2019/0207626A1 (“Ross”) and further in view of Holland, US Pre-Grant Publication 2019/0228284A1 (“Holland”) and Ko, et al., US Pre-Grant Publication 2021/0133570A1 (“Ko”).
Regarding claim 7, Chalfin in view of Ross and Holland teaches the data processing system of claim 3. While Chalfin in view of Ross and Holland teaches a convolutional model that compresses and decompresses model weights per layer, the combination does not explicitly teach wherein the weight decoder selection module selects one of the first weight decoder or the second weight decoder based on a difference between the first characteristic and the second characteristics.
Ko teaches wherein the weight decoder selection module selects one of the first weight decoder or the second weight decoder based on a difference between the first characteristic and the second characteristics. (Ko, ⁋124, “For example, the processor 410 may calculate compression rates corresponding to each of the candidate profiles 61, 62, 63, and 64, and may determine a candidate profile having the highest compression rate among the calculated compression rates as the final profile; determining the final profile, or selection, based taking the highest compression rate is interpreted as comparing based on a difference between a first and second characteristic (i.e. wherein the weight decoder selection module selects one of the first weight decoder or the second weight decoder based on a difference between the first characteristic and the second characteristics.).”).
Chalfin, in view of Ross and Holland, and Ko are both in the same field of endeavor (i.e. model compression). It would have been obvious for a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chalfin, in view of Ross and Holand, and Ko to teach the above limitation(s). The motivation for doing so is that comparing different compression/decompression functions’ compression rates allows for better selection of a compression/decompression function for the given input (cf. Ko, ⁋123, “The processor 410 may determine a final profile by comparing compression performances for each of the generated candidate profiles 61, 62, 63, and 64. The compression performances may include a time complexity of each profile, the size of a required memory space, a compression rate, and statistical characteristics, but are not limited thereto.”).
Regarding claim 10, Chalfin in view of Ross and Holland teaches the data processing system of claim 1. While Chalfin in view of Ross and Holland teaches a convolutional model that compresses and decompresses model weights per layer, the combination does not explicitly teach wherein the data processing system is configured to obtain control data corresponding to the compressed set of weight values, and wherein selecting a weight decoder from the set of weight decoders comprises selecting a weight decoder from the set of weight decoders based on the control data.
Ko teaches wherein the data processing system is configured to obtain control data corresponding to the compressed set of weight values, and wherein selecting a weight decoder from the set of weight decoders comprises selecting a weight decoder from the set of weight decoders based on the control data. (Ko, ⁋182, “may obtain an optimal configuration predetermined by a final profile in which compression techniques set for each lane are defined by dividing neural network data into one or more lanes. For example, the first splitter 142 may obtain the optimal configuration from the memory 520 (refer to FIG. 5), and may obtain from the memory 520 via the neural network processing unit 512 (refer to FIG. 5) a control signal [to obtain control data corresponding to the compressed set of weight values,] generated by the neural network processing unit 512 based on the optimal configuration; getting the final profile from the control signal is interpreted as control data selecting the decoder as the selected compression techniques in the final profile have corresponding decompression techniques (i.e. and wherein selecting a weight decoder from the set of weight decoders comprises selecting a weight decoder from the set of weight decoders based on the control data.).” and Ko, ⁋178, “Meanwhile, the neural network apparatus 5 may compress or decompress neural network data based on an optimal configuration determined in advance through profiling as described above; further support that the profile selection also dictates the decompression.”).
Chalfin, in view of Ross and Holland, and Ko are both in the same field of endeavor (i.e. model compression). It would have been obvious for a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chalfin, in view of Ross and Holland, and Ko to teach the above limitation(s). The motivation for doing so is that final profile data selects the best compression/decompression function for the given input (cf. Ko, ⁋123, “The processor 410 may determine a final profile by comparing compression performances for each of the generated candidate profiles 61, 62, 63, and 64. The compression performances may include a time complexity of each profile, the size of a required memory space, a compression rate, and statistical characteristics, but are not limited thereto.”).
Claims 5-6 and 16-17 are rejected under 35 U.S.C. 103 as being unpatentable over Chalfin, et al., US Pre-Grant Publication 2018/0239992A1 (“Chalfin”) in view of Ross, et al., US Pre-Grant Publication 2019/0207626A1 (“Ross”) and further in view of Holland, US Pre-Grant Publication 2019/0228284A1 (“Holland”) and Pope, et al., US Pre-Grant Publication 2022/0376703A1 (“Pope”).
Regarding claim 5 and analogous claim 16, Chalfin in view of Ross and Holland teaches the data processing system of claim 3. While Chalfin in view of Ross and Holland teaches a convolutional model that compresses and decompresses model weights per layer, the combination does not explicitly teach wherein the first characteristic includes a first measure of computational resources to be used when decompressing the compressed set of weight values using the first weight decoder and the second characteristic includes a second measure of computational resources to be used when decompressing the compressed set of weight values using the second weight decoder.
Pope teaches wherein the first characteristic includes a first measure of computational resources to be used when decompressing the compressed set of weight values using the first weight decoder and the second characteristic includes a second measure of computational resources to be used when decompressing the compressed set of weight values using the second weight decoder. (Pope, ⁋107, “One reason for generating code tables based on groups of code words can be to meet hardware limitations of the system from implementation-to-implementation, for example to generate a code table that can fit in limited memory available to the compressor and decompressor devices; memory available to the decompressor devices is interpreted as a measure of computational resources used when decompressing a set of compressed weight value (i.e. wherein the first characteristic includes a first measure of computational resources to be used when decompressing the compressed set of weight values using the first weight decoder and the second characteristic includes a second measure of computational resources to be used when decompressing the compressed set of weight values using the second weight decoder.).”).
Chalfin, in view of Ross and Holland, and Pope are both in the same field of endeavor (i.e. model compression). It would have been obvious for a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chalfin, in view of Ross and Holland, and Pope to teach the above limitation(s). The motivation for doing so is that considering available resources allows for better selection of compression/decompression configurations (cf. Pope, ⁋107, “In this way, the system can be implemented to balance trade-offs between available hardware resources, such as smaller memory requirements for smaller code tables, versus compression performance realized through higher compression ratios”).
Regarding claim 6 and analogous claim 17, Chalfin in view of Ross and Holland teaches the data processing system of claim 3. While Chalfin in view of Ross and Holland teaches a convolutional model that compresses and decompresses model weights per layer, the combination does not explicitly teach wherein the first characteristic comprises a composite measure of a first compression ratio associated with the first weight decoder and of computational resources to be used when decompressing the compressed set of weight values using the first weight decoder and the second characteristic comprises a composite measure of a second compression ratio associated with the second weight decoder and of computational resources to be used when decompressing the compressed set of weight values using the second weight decoder.
Pope teaches:
wherein the first characteristic comprises a composite measure of a first compression ratio associated with the first weight decoder and of computational resources to be used when decompressing the compressed set of weight values using the first weight decoder (Pope, ⁋107, “One reason for generating code tables based on groups of code words can be to meet hardware limitations of the system from implementation-to-implementation, for example to generate a code table that can fit in limited memory available to the compressor and decompressor devices. In this way, the system can be implemented to balance trade-offs [wherein the first characteristic comprises a composite measure] between available hardware resources, such as smaller memory requirements for smaller code tables, versus compression performance realized through higher compression ratios for larger code tables [and of computational resources to be used when decompressing the compressed set of weight values using the first weight decoder].”).
and the second characteristic comprises a composite measure of a second compression ratio associated with the second weight decoder and of computational resources to be used when decompressing the compressed set of weight values using the second weight decoder. (Pope, ⁋107, “One reason for generating code tables based on groups of code words can be to meet hardware limitations of the system from implementation-to-implementation, for example to generate a code table that can fit in limited memory available to the compressor and decompressor devices. In this way, the system can be implemented to balance trade-offs [and the second characteristic comprises a composite measure] between available hardware resources, such as smaller memory requirements for smaller code tables, versus compression performance realized through higher compression ratios for larger code tables [of a second compression ratio associated with the second weight decoder and of computational resources to be used when decompressing the compressed set of weight values using the second weight decoder.].”).
Chalfin, in view of Ross and Holland, and Pope are both in the same field of endeavor (i.e. model compression). It would have been obvious for a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Chalfin, in view of Ross and Holland, and Pope to teach the above limitation(s). The motivation for doing so is that considering available resources and compression ratios allows for better selection of compression/decompression configurations (cf. Pope, ⁋107, “In this way, the system can be implemented to balance trade-offs between available hardware resources, such as smaller memory requirements for smaller code tables, versus compression performance realized through higher compression ratios”).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Lee, et al., US20210111741A1 discloses using a decompression method that quantizes weight matrixes and then dequantizes the weight matrix on inference time.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to NICHOLAS S WU whose telephone number is (571)270-0939. The examiner can normally be reached Monday - Friday 8:00 am - 4:00 pm EST.
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, Michelle Bechtold can be reached at 571-431-0762. 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.
/N.S.W./Examiner, Art Unit 2148 /MICHELLE T BECHTOLD/Supervisory Patent Examiner, Art Unit 2148