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 .
Amendments
This Office Action is in response to the amendment filed on January 16, 2026.
Claim(s) 26, 33, 38, and 42 have been amended.
Claims 27, 39, and 43 have been cancelled.
No new claims have been added.
The objections and rejections from the prior correspondence that are not restated herein are withdrawn.
Response to Arguments
Applicant's arguments filed on January 16, 2026 have been fully considered.
Applicant's arguments regarding the 35 U.S.C. 101 rejections of the previous office action have been fully considered but are not persuasive. Applicant argues: “Applicant respectfully disagrees and respectfully submits that the claims do not recite an abstract idea. […] No specific mathematical calculation is recited, nor is a method of organizing human activity recited.”
Examiner respectfully disagrees. Independent claims 26 and 33 recite the abstract idea of decoding sizes, ranks, and tensors from a bitstream. Page 33 of the specification of the instant application describes that decoding encompasses processes such as inverse quantization, inverse transformation, and differential decoding. Therefore, the broadest reasonable interpretation of claims 26 and 33 encompass mathematical calculations for decoding sizes, ranks, and tensors, as described in the specification. Additionally, independent claims 38 and 42 recite the abstract idea of decomposing a first tensor into a second and third tensor. Page 6 of the specification of the instant application describes determining the low-rank representation of a matrix (i.e., tensor) using single value decomposition (SVD). Therefore, the broadest reasonable interpretation of decomposing a [first] tensor encompasses SVD mathematical calculations. Furthermore, claims 38 and 42 recite deriving one or more decomposition ranks. Page 22 of the specification describes applying mean squared error (MSE) thresholds to derive decomposition ranks. For this reason, the independent claims recite abstract ideas that encompass mathematical calculations.
Applicant argues: “Additionally, the claims cannot be performed in the human mind as tensors are decomposed at a Deep Neural Network. For example, the subject-matter of the claims including decoding and encoding using two tensors may be enhanced by utilizing or signaling a dimension of a tensor and a decomposition rank of a tensor. For example, a size of a tensor may be used to encode or decode which may result in more efficient encoding and decoding since the dimension of each tensor is known. Additionally, a decomposition rank of a tensor may be used to encode or decode which may result in more efficient encoding and decoding since the decomposition rank of each tensor is known.”
Examiner respectfully disagrees. It should be noted that only the abstract limitations of claims 31 and 36 are being rejected as mental processes. As noted above, decomposition of tensors encompasses mathematical calculations. Additionally, reciting that the decomposition is performed at a Deep Neural Network would be considered using a neural network as a tool for implementing the abstract idea of decomposing a tensor. Regarding limitations performed in the human mind, claims 31 and 36 recite determining/determine if one or more of the decoded second tensor and the decoded third tensor is in the decoded tensor buffer by looking for a tensor associated with an identifier. The broadest reasonable interpretation for determine […] by looking are within the capabilities of a mental evaluation/observation of a person. Thus, claim 31 and 36 recite limitations that can be performed in the human mind.
Applicant's arguments regarding the 35 U.S.C. 103 rejections of the previous office action have been fully considered but are not persuasive. Applicant argues: “Minezawa discusses using quantization information which is decoded from compressed data. However, the quantization information appears to be a weight of a parameter (see Minezawa at paragraphs 0044, 0045, and 0054). There is no suggestion in Minezawa to decode one or more decomposition ranks corresponding to at least one or more of the second tensor and the third tensor nor to decode the second tensor and the third tensor based on the one or more decomposition ranks. Sainath does not discuss any decoding of a tensor based on one or more decomposition ranks either. Instead, Sainath merely shares a low-rank projection matrix across two layers (see Sainath at paragraph 0037).”
Examiner respectfully disagrees. It should be noted that the combination of SAINATH, MINEZAWA, and CRICI teaches the recited limitations for claims 26 (and 33 accordingly) as shown in detail below:
decoding from the bitstream one or more decomposition ranks corresponding to at least one or more of the second tensor and the third tensor; (CRICI [0109-0110] teaches performing "Matrix Decomposition" as a step in a compression pipeline. CRICI [0123] teaches that "decoding_step_id" identifies the decoding process or step to be performed. CRICI [0175] teaches: "Tensor Dimensions. tensor_dimensions may be a field of the Layer Parameter Set and may specify the dimensions of the tensor to which the layer parameter set refers. In another embodiment, tensor_dimensions may be a field of the NNR compressed data unit’s header and may specify the dimensions of the tensor (i.e., one or more decomposition ranks corresponding to at least one or more of the second tensor and the third tensor) carried in the payload of the same NNR compressed data unit." CRICI [0061] teaches: "A bitstream is formed by concatenating several NNR Units. NNR Units may contain different types of data. The type of data that is contained in the payload of an NNR Unit defines the NNR Unit’s type." CRICI [0109-0110] and [0175-0176] teaches the pipeline that applies quantization and entropy coding to the decomposed tensor dimensions carried in the payload in a bitstream (i.e., decoding from the bitstream).)
and decoding the second tensor and the third tensor […] based on the one or more decomposition ranks (CRICI [0057] teaches: “The decoder 505 uses a decoder or decompression algorithm, for example, to perform the neural network decoding 506 to decode the compressed data 509 (for example, compressed video) which was encoded by the encoder 503. The decoder 505 produces decompressed data 510 (for example, reconstructed data) (i.e., the second tensor and the third tensor).” CRICI [0174] teaches: “data_size may indicate the number of parameters or weights belong to this id when the compressed N data unit is uncompressed. In another embodiment, this value may indicate the byte size which corresponds to such parameters or weights.” CRICI [0175] teaches: “Extension 9: Tensor Dimensions. tensor_dimensions may be a field of the Layer Parameter Set and may specify the dimensions of the tensor to which the layer parameter set refers.” CRICI [0192] teaches: “In an embodiment, ONNX message identifier types may be signalled in the corresponding NNR unit headers so that corresponding NNR unit payloads could be parsed and processed correctly.” CRICI [0061] teaches: “A bitstream is formed by concatenating several NNR Units.” Examiner’s note: NNR stands for Neural Network Representation (see CRICI [0024]). Under broadest reasonable interpretation, the one or more decomposition ranks can be interpreted as the tensor dimensions field for the several NNR units in the compressed data unit’s header. Additionally, the second tensor and third tensor can be interpreted as the reconstructed data from several NNR units. Therefore, decoding the compressed data, which includes the tensor dimensions, teaches decoding […] based on based on the one or more decomposition ranks to obtain the decoded second tensor and a decoded third tensor.)
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 26, 28-38, 40-42, and 44-45 are rejected under 35 U.S.C.101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1: Claims 26, 28-32, 38, and 40-41 are directed to a process. Claims 33-37, 42 and 44-45 are directed to a machine or an article of manufacture.
With respect to claim(s) 26 and 33:
2A Prong 1: The claim recites an abstract idea. Specifically:
decoding from the bitstream one or more sizes corresponding to at least one or more of the second tensor and the third tensor (Mathematical calculations – see MPEP § 2106.04(a)(2)(I))
and decoding from the bitstream one or more decomposition ranks corresponding to at least one or more of the second tensor and the third tensor; (Mathematical concepts – Decoding involves mathematical calculations (see pg. 33, lines 33-36) – see MPEP § 2106.04(a)(2)(I))
decoding the second tensor and the third tensor based on the one or more decoded sizes and based on the one or more decomposition ranks to obtain the decoded second tensor and a decoded third tensor. (Mathematical concepts – Decoding involves mathematical calculations (see pg. 33, lines 33-36) – see MPEP § 2106.04(a)(2)(I))
2A Prong 2: The additional elements recited in the claim do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
(Claim 33) An apparatus comprising one or more processors, the one or more processors configured to: (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
responsive to a determination that a first tensor of a layer of a Deep Neural Network is decomposed into a second tensor and a third tensor whose parameters are encoded in a bitstream (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
(Claim 33) An apparatus comprising one or more processors, the one or more processors configured to: (Mere instructions to apply an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
responsive to a determination that a first tensor of a layer of a Deep Neural Network is decomposed into a second tensor and a third tensor whose parameters are encoded in a bitstream (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
Therefore, the claim is ineligible.
With respect to claim(s) 28 and 34:
2A Prong 1: The claim recites an abstract idea. Specifically:
deriving one or more sizes of one or more of the second tensor or the third tensor based on the one or more decoded sizes (Mathematical calculations – see MPEP § 2106.04(a)(2)(I))
decoding one or more of the second tensor and the third tensor based on the one or more derived sizes (Mathematical calculations – see MPEP § 2106.04(a)(2)(I)).
Additionally, the claim does not recite any new additional elements that would amount to an integration of the abstract idea into a practical application (individually or in combination) or significantly more than the judicial exception.
Therefore, the claim is ineligible.
With respect to claim(s) 29 and 35:
2A Prong 1: The claim recites an abstract idea. Specifically:
reconstructing the first tensor based on the decoded second tensor and the decoded third tensor (Mathematical calculations – see MPEP § 2106.04(a)(2)(I)).
Additionally, the claim does not recite any new additional elements that would amount to an integration of the abstract idea into a practical application (individually or in combination) or significantly more than the judicial exception.
Therefore, the claim is ineligible.
With respect to claim(s) 30, 40, 44:
2A Prong 2: The additional elements recited in the claim do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
(Claim 44) one or more processors (Mere recitation of a generic computer component – see § MPEP 2106.05(b)(I))
storing one or more of the decoded second tensor and the decoded third tensor in a decoded tensor buffer (Insignificant extra solution activity of mere data gathering – see § MPEP2106.05(g).)
2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
(Claim 44) one or more processors (Mere recitation of a generic computer component – see § MPEP 2106.05(b)(I))
storing one or more of the decoded second tensor and the decoded third tensor in a decoded tensor buffer (Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception (WURC)- see MPEP § 2106.05(d)(ll)(iv) - Storing and retrieving information in memory, Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93.)
Therefore, the claim is ineligible.
With respect to claim(s) 31:
2A Prong 1: The claim recites an abstract idea. Specifically:
determining if one or more of the decoded second tensor and the decoded third tensor is in the decoded tensor buffer by looking for a tensor associated with an identifier (Mental process – see MPEP § 2106.04(a)(2)(III))
2A Prong 2: The additional elements recited in the claim do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
the identifier comprising a same layer as the one or more of the decoded second tensor and the decoded third tensor (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
the identifier comprising a same layer as the one or more of the decoded second tensor and the decoded third tensor (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
Therefore, the claim is ineligible.
With respect to claim(s) 32 and 37:
2A Prong 2: The additional elements recited in the claim do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
wherein the bitstream includes additional parameters associated with the first tensor (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
wherein the bitstream includes additional parameters associated with the first tensor (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
Therefore, the claim is ineligible.
With respect to claim(s) 36:
2A Prong 1: The claim recites an abstract idea. Specifically:
determine if one or more of the decoded second tensor and the decoded third tensor is in the decoded tensor buffer by looking for a tensor associated with an identifier (Mental process – see MPEP § 2106.04(a)(2)(III))
2A Prong 2: The additional elements recited in the claim do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
store one or more of the decoded second tensor and the decoded third tensor in a decoded tensor buffer (Insignificant extra solution activity of mere data gathering – see § MPEP2106.05(g).)
2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
store one or more of the decoded second tensor and the decoded third tensor in a decoded tensor buffer (Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception (WURC)- see MPEP § 2106.05(d)(ll)(iv) - Storing and retrieving information in memory, Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015); OIP Techs., 788 F.3d at 1363, 115 USPQ2d at 1092-93.)
Therefore, the claim is ineligible.
With respect to claim(s) 38 and 42:
2A Prong 1: The claim recites an abstract idea. Specifically:
decomposing a first tensor of a layer of a Deep Neural Network into a second tensor and a third tensor; (Mathematical calculations – see MPEP § 2106.04(a)(2)(I))
deriving one or more sizes corresponding to at least one or more of the second tensor and the third tensor; (Mathematical calculations – see MPEP § 2106.04(a)(2)(I))
deriving one or more decomposition ranks corresponding to at least one or more of the second tensor and the third tensor; (Mathematical concepts – Deriving decomposition ranks is performed using mean squared error (MSE) thresholds (see pg. 22, lines 1-19) – see MPEP § 2106.04(a)(2)(I))
encoding the second tensor and the third tensor in a bitstream based on the determined one or more sizes and based on the one or more decomposition ranks (Mathematical concepts – Decomposition ranks are derived using mean squared error (MSE) thresholds (see pg. 22, lines 1-19) – see MPEP § 2106.04(a)(2)(I))
2A Prong 2: The additional elements recited in the claim do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
(Claim 42) one or more processors (Mere recitation of a generic computer component – see § MPEP 2106.05(b)(I))
wherein the one or more sizes corresponding to at least one or more of the second tensor and the third tensor are encoded in the bitstream (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
(Claim 42) one or more processors (Mere recitation of a generic computer component – see § MPEP 2106.05(b)(I))
wherein the one or more sizes corresponding to at least one or more of the second tensor and the third tensor are encoded in the bitstream (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea – see MPEP 2106.05(f).)
Therefore, the claim is ineligible.
With respect to claim(s) 41 and 45:
2A Prong 2: The additional elements recited in the claim do not integrate the abstract idea into a practical application, individually or in combination.
Additional elements:
(Claim 45) one or more processors (Mere recitation of a generic computer component – see § MPEP 2106.05(b)(I))
transmitting the bitstream to a decoder (Adding insignificant extra-solution activity to the judicial exception – see § MPEP2106.05(g).)
2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
(Claim 45) one or more processors (Mere recitation of a generic computer component – see § MPEP 2106.05(b)(I))
transmitting the bitstream to a decoder (Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception (WURC)- see MPEP § 2106.05(d)(ll)(i) - Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information).)
Therefore, the claim is ineligible.
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 26, 29-33, and 35-37 are rejected under 35 U.S.C. 103 as being unpatentable over SAINATH (US 20170076196 A1) in view of MINEZAWA (US 20200184318 A1) and CRICI (US 20230209092 A1), hereafter SAINATH, MINEZAWA, and CRICI respectively.
Regarding Claim 26:
SAINATH teaches:
[…] a first tensor of a layer of a Deep Neural Network is decomposed into a second tensor and a third tensor […] (SAINATH [0026] and [0037] teaches an uncompressed recurrent parameter matrix m*n (i.e., a first tensor) of a deep LSTM layer (i.e., a layer of a Deep Neural Network) being compressed by replacing (i.e., decomposed) the uncompressed recurrent parameter matrix with a compressed recurrent parameter matrix m*r (i.e., second tensor) and a projection matrix r*n (i.e., third tensor).)
SAINATH is not relied upon for teaching:
responsive to a determination that […] whose parameters are encoded in a bitstream, decoding from the bitstream one or more sizes corresponding to at least one or more of the second tensor and the third tensor, and decoding from the bitstream one or more decomposition ranks corresponding to at least one or more of the second tensor and the third tensor; and
decoding the second tensor and the third tensor based on the one or more decoded sizes and based on the one or more decomposition ranks to obtain the decoded second tensor and a decoded third tensor.
However, MINEZAWA teaches: responsive to a determination that a first tensor of a layer of a Deep Neural Network is decomposed into a second tensor and a third tensor decoding from the bitstream one or more sizes corresponding to at least one or more of the second tensor and the third tensor; (MINEZAWA [0040] and [0053] teaches decoding neural network configuration information from the compressed data. Some of the information included in the neural network configuration information is the number of network layers, number of nodes for each of the layers, and edges that link nodes. This information describes the layer configuration for the neural network in the compressed data. Furthermore, one could recognize that applying the data compression from MINEZAWA to matrices from SAINATH can generate the compressed data, and the decoding is the result of (i.e., responsive to) receiving the compressed data for decoding. Moreover, the one or more sizes can be understood as the dimensions of the parameter matrix m*r and projection matrix r*n that are encoded in the compressed data (e.g. m, n, and r are sizes or dimensions of a tensor).)
[…] whose parameters are encoded in a bitstream, decoding the second tensor and the third tensor based on the one or more decoded sizes to obtain the decoded second tensor and a decoded third tensor. (MINEZAWA [0050] teaches encoding quantization information and neural network configuration information to generate compressed data (i.e., a bitstream). The parameters of the compressed recurrent parameter matrix (i.e., second tensor) and projection matrix (i.e., third tensor) correspond to the neural network configuration information encoded in the compressed data. MINEZAWA [0053], [0095], and [0172-0175] teaches decoding compressed data and inversely quantizing the quantization information and neural network configuration information to construct the neural network on the decoding side. Some of the information included in the neural network configuration information is the number of network layers, number of nodes for each of the layers, and edges that link nodes.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of MINEZAWA and SAINATH before them, to include MINEZAWA's neural network configuration information encoding and compressed data decoding into SAINATH's neural network compression method. One would have been motivated to make such a combination in order to optimize a neural network on the encoding side for construction on the decoding side (MINEZAWA [0015]).
SAINATH in view of MINEZAWA is not relied upon for teaching, but CRICI teaches: and decoding from the bitstream one or more decomposition ranks corresponding to at least one or more of the second tensor and the third tensor; and (CRICI [0109-0110] teaches performing "Matrix Decomposition" as a step in a compression pipeline. CRICI [0123] teaches that "decoding_step_id" identifies the decoding process or step to be performed. CRICI [0175] teaches: "Tensor Dimensions. tensor_dimensions may be a field of the Layer Parameter Set and may specify the dimensions of the tensor to which the layer parameter set refers. In another embodiment, tensor_dimensions may be a field of the NNR compressed data unit’s header and may specify the dimensions of the tensor (i.e., one or more decomposition ranks corresponding to at least one or more of the second tensor and the third tensor) carried in the payload of the same NNR compressed data unit." CRICI [0061] teaches: "A bitstream is formed by concatenating several NNR Units. NNR Units may contain different types of data. The type of data that is contained in the payload of an NNR Unit defines the NNR Unit’s type." CRICI [0109-0110] and [0175-0176] teaches the pipeline that applies quantization and entropy coding to the decomposed tensor dimensions carried in the payload in a bitstream (i.e., decoding from the bitstream).)
decoding the second tensor and the third tensor […] and based on the one or more decomposition ranks to obtain the decoded second tensor and a decoded third tensor. (CRICI [0057] teaches: “The decoder 505 uses a decoder or decompression algorithm, for example, to perform the neural network decoding 506 to decode the compressed data 509 (for example, compressed video) which was encoded by the encoder 503. The decoder 505 produces decompressed data 510 (for example, reconstructed data) (i.e., the second tensor and the third tensor).” CRICI [0174] teaches: “data_size may indicate the number of parameters or weights belong to this id when the compressed N data unit is uncompressed. In another embodiment, this value may indicate the byte size which corresponds to such parameters or weights.” CRICI [0175] teaches: “Extension 9: Tensor Dimensions. tensor_dimensions may be a field of the Layer Parameter Set and may specify the dimensions of the tensor to which the layer parameter set refers.” CRICI [0192] teaches: “In an embodiment, ONNX message identifier types may be signalled in the corresponding NNR unit headers so that corresponding NNR unit payloads could be parsed and processed correctly.” CRICI [0061] teaches: “A bitstream is formed by concatenating several NNR Units.” Examiner’s note: NNR stands for Neural Network Representation (see CRICI [0024]). Under broadest reasonable interpretation, the one or more decomposition ranks can be interpreted as the tensor dimensions field for the several NNR units in the compressed data unit’s header. Additionally, the second tensor and third tensor can be interpreted as the reconstructed data from several NNR units. Therefore, decoding the compressed data, which includes the tensor dimensions, teaches decoding […] based on based on the one or more decomposition ranks to obtain the decoded second tensor and a decoded third tensor.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of SAINATH, MINEZAWA and CRICI before them, to include CRICI’s payload decoding tensor dimensions in SAINATH and MINEZAWA's neural network compression method. One would have been motivated to make such a combination in order to address many missing aspects of an interoperable information exchange mechanism, carriage of configuration and parameters related to the exchanged data of compressed neural networks (CRICI [0086]).
Regarding Claim 29:
SAINATH in view of MINEZAWA and CRICI teaches the elements of claim 26 as outlined above. MINEZAWA further teaches:
further comprising reconstructing the first tensor based on the decoded second tensor and the decoded third tensor. (MINEZAWA [0015] teaches constructing (i.e., reconstructing) a neural network (i.e., the first tensor) on the decoding side based on the quantization information and network configuration information decoded from the compressed data.)
Regarding Claim 30:
SAINATH in view of MINEZAWA and CRICI teaches the elements of claim 26 as outlined above. MINEZAWA further teaches:
The storing one or more of the decoded second tensor and the decoded third tensor in a decoded tensor buffer. (MINEZAWA [0059], [0079], [0082], [0093-0096], and [0119] teaches decoding from the compressed data and outputting the decoding results from the decoding unit 201 to the data processing unit 202. The processor 301 implements the functions of the decoding unit 201 and data processing unit 202, and constructs the neural network using the decoded data. Additionally, Fig. 3B shows the processor 301 and the memory 302 connected to each other by a signal bus, which transmits the data. A person having ordinary skill in the art would recognize that while neural network construction occurs, the decoded data required for the neural network construction must be stored in memory (i.e., decoded tensor buffer), because constructing the neural network requires a large memory size as disclosed in MINEZAWA.)
Regarding Claim 31:
SAINATH in view of MINEZAWA and CRICI teaches the elements of claim 30 as outlined above. CRICI further teaches:
determining if one or more of the decoded second tensor and the decoded third tensor is in the decoded tensor buffer by looking for a tensor associated with an identifier, the identifier comprising a same layer as the one or more of the decoded second tensor and the decoded third tensor. (CRICI [0160], [0166-0167], and [0171]-[0186] teaches identifying a compressed NN data unit, such as a tensor, in the same layer by providing an id_name unique identifier (i.e., associated with an identifier) in the Layer Parameter Set's lps_id_list. It uses context_mode and context_id to select previously decoded symbols as the context for decoding, which tells the decoder whether the current tensor being processed has already been decoded. The decoded second and third tensors in the tensor buffer are taught by MINEZAWA [0059], [0079], [0082], [0093-0096], and [0119] as outlined in claim 30 above.)
Regarding Claim 32:
SAINATH in view of MINEZAWA and CRICI teaches the elements of claim 26 as outlined above. MINEZAWA further teaches:
The method of claim 26, wherein the bitstream includes additional parameters associated with the first tensor. (MINEZAWA [0015] teaches additional parameters associated with the first tensor as the quantization information, which is encoded along with the network configuration in the compressed data (i.e., bitstream). The first tensor is taught by SAINATH [0026] and [0037] as outlined above in claim 26.)
Regarding Claim 33:
The claim recites similar limitations as corresponding claim 26 and is rejected for similar reasons as claim 26 using similar teachings and rationale. Additionally, MINEZAWA teaches:
one or more processors (MINEZAWA [0067] teaches processor 301.)
Regarding Claim 35:
SAINATH in view of MINEZAWA and CRICI teaches the elements of claim 33 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 29 and is rejected for similar reasons as claim 29 using similar teachings and rationale.
Regarding Claim 36:
SAINATH in view of MINEZAWA and CRICI teaches the elements of claim 33 as outlined above. Additionally, the claim recites similar limitations as corresponding claims 30 and 31 and is rejected for similar reasons as claims 30 and 31 using similar teachings and rationale.
Regarding Claim 37:
SAINATH in view of MINEZAWA and CRICI teaches the elements of claim 33 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 32 and is rejected for similar reasons as claim 32 using similar teachings and rationale.
Claims 28 and 34 are rejected under 35 U.S.C. 103 as being unpatentable over SAINATH in view of MINEZAWA and CRICI as applied to claims 26 and 33 respectively above, and further in view of CHOE (KR 20200064348 A), hereafter CHOE.
Regarding Claim 28:
SAINATH in view of MINEZAWA and CRICI teaches the elements of claim 26 as outlined above. MINEZAWA further teaches:
[…] one or more decoded sizes; (MINEZAWA [0040] and [0053] teaches decoding neural network configuration information from the compressed data as outlined in claim 26 above.)
decoding one or more of the second tensor and the third tensor based on the one or more […] sizes. (MINEZAWA [0053], [0095], and [0172-0175] teaches decoding compressed data and inversely quantizing the quantization information and neural network configuration information to construct the neural network on the decoding side.)
SAINATH in view of CRICI and MINEZAWA is not relied upon for teaching but CHOE teaches: deriving one or more sizes of one or more of the second tensor or the third tensor […] (CHOE [33], [40] and equation (6) teaches calculating (i.e., deriving) a tensor rank by using the tensor width, tensor height (i.e., one or more sizes), and number of dimensions of the input tensor (i.e., one or more […] tensor). A person having ordinary skill in the art could apply the rank calculation from CHOE to the decomposed tensors of the combination of SAINATH, CRICI, and MINEZAWA.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of SAINATH, MINEZAWA, CRICI and CHOE before them, to include CHOE's rank calculation in SAINATH, MINEZAWA, and CRICI’s neural network information compression method. One would have been motivated to make such a combination because by applying the first rank and the second rank determined through the lower limit, it is possible to minimize the time required in the compression process while providing the same compression rate (CHOE [43]).
Regarding Claim 34:
SAINATH in view of MINEZAWA and CRICI teaches the elements of claim 33 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 28 and is rejected for similar reasons as claim 28 using similar teachings and rationale.
Claims 38, 40-42, and 44-45 are rejected under 35 U.S.C. 103 as being unpatentable over SAINATH in view of CHOE, MINEZAWA, and CRICI.
Regarding Claim 38:
SAINATH teaches:
decomposing a first tensor of a layer of a Deep Neural Network into a second tensor and a third tensor; (SAINATH [0026] and [0037] teaches an uncompressed recurrent parameter matrix m*n (i.e., a first tensor) of a deep LSTM layer (i.e., layer of a Deep Neural Network) being compressed by replacing (i.e., decompose) the uncompressed recurrent parameter matrix with a compressed recurrent parameter matrix m*r (i.e., second tensor) and a projection matrix r*n (i.e., third tensor).
SAINATH is not relied upon for teaching:
deriving one or more sizes corresponding to at least one or more of the second tensor and the third tensor; and
deriving one or more decomposition ranks corresponding to at least one or more of the second tensor and the third tensor;
encoding the second tensor and the third tensor in a bitstream based on the determined one or more sizes and based on the one or more decomposition ranks, wherein the one or more sizes corresponding to at least one or more of the second tensor and the third tensor are encoded in the bitstream.
However, CHOE teaches: deriving one or more sizes corresponding to at least one or more of the second tensor and the third tensor; (CHOE [33], [40] and equation (6) teaches calculating (i.e., deriving) a tensor rank by using the tensor width, tensor height (i.e., one or more sizes), and number of dimensions of the input tensor.)
deriving one or more decomposition ranks corresponding to at least one or more of the second tensor and the third tensor; (CHOE [33], [40] and equation (6) teaches calculating a tensor rank (i.e., deriving one or more decomposition ranks) by using the tensor width, tensor height, and number of dimensions of the input tensor.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of SAINATH and CHOE before them, to include CHOE's rank calculation in SAINATH’s neural network information compression method. One would have been motivated to make such a combination because by applying the first rank and the second rank determined through the lower limit, it is possible to minimize the time required in the compression process while providing the same compression rate (CHOE [43]).
SAINATH in view of CHOE is not relied upon for teaching:
encoding the second tensor and the third tensor in a bitstream based on the determined one or more sizes and based on the one or more decomposition ranks, wherein the one or more sizes corresponding to at least one or more of the second tensor and the third tensor are encoded in the bitstream.
However, MINEZAWA teaches: wherein the one or more sizes corresponding to at least one or more of the second tensor and the third tensor are encoded in the bitstream. (MINEZAWA [0049] teaches encoding the network configuration information to generate compressed data (i.e., bitstream). SAINATH teaches the decomposed neural network layer matrix m*n (i.e., a first tensor) into m*r (i.e., second tensor) and r*n (i.e., third tensor), which correspond to the sizes of each matrix respectively. Therefore, the combination of MINEZAWA and SAINATH teaches encoding of neural network configuration information to generate compressed data.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of SAINATH, CHOE, and MINEZAWA before them, to include MINEZAWA's neural network configuration information encoding into SAINATH and CHOE’s neural network compression method. One would have been motivated to make such a combination in order to optimize a neural network on the encoding side for construction on the decoding side (MINEZAWA [0015]).
SAINATH in view of CHOE and MINEZAWA is not relied upon for teaching, but CRICI teaches: encoding the second tensor and the third tensor in a bitstream based on the determined one or more size and based on the one or more decomposition ranks (CRICRI [0056] teaches: “The encoder 503 has the goal of compressing input data 507 (for example, an input video) to compressed data 509 (for example, a bitstream) (i.e., encoding the second and the third tensor in a bitstream) […].” CRICI [0175] teaches: “In another embodiment, tensor_dimensions may be a field of the NNR compressed data unit’s header and may specify the dimensions of the tensor (i.e., based on the determined one or more sizes and based on the one or more decomposition ranks) carried in the payload of the same NNR compressed data unit.” CRICI [0061] teaches: “A bitstream is formed by concatenating several NNR Units.” Examiner’s note: A rank and size are tensor dimensions.)
Accordingly, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention, having the teachings of SAINATH, CHOE, MINEZAWA and CRICI before them, to include CRICI’s payload encoding tensor dimensions in SAINATH, CHOE, and MINEZAWA's neural network compression method. One would have been motivated to make such a combination in order to address many missing aspects of an interoperable information exchange mechanism, carriage of configuration and parameters related to the exchanged data of compressed neural networks (CRICI [0086]).
Regarding Claim 40:
SAINATH in view of CHOE, MINEZAWA, and CRICI teaches the elements of claim 38 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 30 and is rejected for similar reasons as claim 30 using similar teachings and rationale.
Regarding Claim 41:
SAINATH in view of CHOE, MINEZAWA, and CRICI teaches the elements of claim 38 as outlined above. MINEZAWA further teaches:
transmitting the bitstream to a decoder (MINEZAWA [Fig. 1] and [Fig. 2] teaches the data processing device 100 that includes an encoding unit 103 that sends (i.e., transmitting) the compressed data (i.e., a bitstream) to the decoding unit 201 (i.e., to a decoder).)
Regarding Claim 42:
The claim recites similar limitations as corresponding claim 38 and is rejected for similar reasons as claim 38 using similar teachings and rationale.
Regarding Claim 44:
SAINATH in view of CHOE, MINEZAWA, and CRICI teaches the elements of claim 42 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 30 and is rejected for similar reasons as claim 30 using similar teachings and rationale.
Regarding Claim 45:
SAINATH in view of CHOE, MINEZAWA, and CRICI teaches the elements of claim 42 as outlined above. Additionally, the claim recites similar limitations as corresponding claim 41 and is rejected for similar reasons as claim 41 using similar teachings and rationale.
Conclusion
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 Alvaro S Laham Bauzo whose telephone number is (571)272-5650. The examiner can normally be reached Mon-Fri 7:30 AM - 11:00 AM | 1:00 PM - 5:30 PM ET.
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, Usmaan Saeed can be reached on (571) 272-4046. 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.
/A.S.L./Examiner, Art Unit 2146
/USMAAN SAEED/Supervisory Patent Examiner, Art Unit 2146