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 .
DETAILED ACTION
Claims 1-19 are present in this application. Claims 1-19 are pending in this office
action.
This office action is NON-FINAL.
Drawings
The Drawings filed on 06/28/23 are acceptable for examination purposes.
Specification
5, The Specification filed on 06/28/23 is acceptable for examination purposes.
Information Disclosure Statement
The information disclosure statements (IDS) filed on 06/28/23 has
been considered by the Examiner and made of record in the application file.
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 1-19 are rejected under 35 U.S.C. 101 because the claimed invention is
directed to an abstract idea without significantly more.
Claim 1 recites, “receiving weights of respective clients, wherein each weight has a respectively corresponding precision that is initially an inherent precision; using a dequantizer to change the weights such that the precisions thereof are changed from the inherent precisions to a same reference precision; determining masks respectively corresponding to the weights based on the inherent precisions; based on the masks, determining an integrated weight by merging the weights having the reference precision; and quantizing the integrated weight to generate quantized weights having the inherent precisions, respectively, and transmitting the quantized weights to the clients.”.
The limitation of “receiving weights of respective clients, wherein each weight has a respectively corresponding precision that is initially an inherent precision; using a dequantizer to change the weights such that the precisions thereof are changed from the inherent precisions to a same reference precision; determining masks respectively corresponding to the weights based on the inherent precisions; based on the masks, determining an integrated weight by merging the weights having the reference precision; and quantizing the integrated weight to generate quantized weights having the inherent precisions, respectively, and transmitting the quantized weights to the clients”. Nothing in the claim element precludes the step from practically
being performed in the mind. Accordingly, the claim recites an abstract idea. This
judicial exception is not integrated into a practical application. The claim is directed to
an abstract idea. The claim does not include additional elements that are sufficient to
amount to significantly more than the judicial exception. The claim is not patent eligible.
Claim 2 is dependent on claim 1 and includes all the limitations of claim 1. Claim
2 recites wherein the dequantizer comprises: blocks, wherein each of the blocks has an input precision and an output precision in claim 2. But each of the blocks has an input precision and an output precision does not go beyond the abstract idea itself.
There are no additional components in the claim that would make it significantly more
than the abstract idea.
Claim 3 is dependent on claim 2 and includes all the limitations of claim 2. Claim
3 recites wherein the changing comprises: inputting each of the weights to whichever of the blocks has an input precision that matches its inherent precision; and obtaining an output of whichever of the blocks has an output precision that matches the reference precision in claim 3. But each of the weights to whichever of the blocks has an input precision that matches its inherent precision does not go beyond the abstract idea itself.
There are no additional components in the claim that would make it significantly more
than the abstract idea.
Claim 4 is dependent on claim 1 and includes all the limitations of claim 1. Claim
4 recites wherein the determining of the masks comprises: obtaining a statistical value of first weights, among the weights, which have an inherent precision greater than or equal to a preset threshold precision among the weights; and determining the masks based on the statistical value in claim 4. But obtaining a statistical value of first weights, among the weights does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 5 is dependent on claim 4 and includes all the limitations of claim 4. Claim
5 recites wherein the determining of the masks based on the statistical value comprises: for each of second weights of which an inherent precision is less than the statistical value among the weights, obtaining a similarity thereof to the statistical value; and determining masks respectively corresponding to the second weights based on the similarities in claim 5. But obtaining a similarity thereof to the statistical value; and determining masks respectively corresponding to the second weights based on the similarities does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 6 is dependent on claim 5 and includes all the limitations of claim 5. Claim
6 recites wherein the determining of the masks respectively corresponding to the second weights comprises: determining a binary mask that maximizes the similarities of the respective second weights in claim 6. But determining a binary mask that maximizes the similarities of the respective second weights does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 7 is dependent on claim 1 and includes all the limitations of claim 1. Claim
6 recites training the dequantizer on a periodic basis in claim 7. But training the dequantizer on a periodic basis does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 8 is dependent on claim 7 and includes all the limitations of claim 7. Claim
8 recites wherein the dequantizer comprises: blocks, wherein the training of the dequantizer comprises: receiving learning weight data; generating pieces of quantized weight data by quantizing the learning weight data; obtaining, for each of the blocks, a first loss that is determined based on a difference between intermediate output weight data predicted from a block and quantized weight data corresponding to the block; obtaining a second loss that is determined based on a difference between final output weight data output from the dequantizer receiving the learning weight data and true weight data corresponding to the learning weight data; and training the dequantizer based on the first loss and the second loss.in claim 8. But generating pieces of quantized weight data by quantizing the learning weight data does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 9 is dependent on claim 1 and includes all the limitations of claim 1. Claim
9 recites training the dequantizer on a periodic basis in claim 9. But training the dequantizer on a periodic basis does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 10 is dependent on claim 1 and includes all the limitations of claim 1. Claim 10 recites training the dequantizer on a periodic basis in claim 10. But training the dequantizer on a periodic basis does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 11 recites the same limitations as claim 1 above. Therefore, claim
11 is rejected based on the same reasoning.
Claim 12 is dependent on claim 11 and includes all the limitations of claim 11. Claim 12 recites wherein each of the blocks has an input precision corresponding
to at least one of the inherent precisions and has an output precision corresponding to at least one of the inherent precisions in claim 12. But wherein each of the blocks has an input precision corresponding to at least one of the inherent precisions and has an output precision corresponding to at least one of the inherent precisions does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 13 recites the same limitations as claim 3 above. Therefore, claim
13 is rejected based on the same reasoning.
Claim 14 is dependent on claim 11 and includes all the limitations of claim 11. Claim 14 recites obtain a statistical value of first weights selected from among the weights based on having an inherent precision greater than or equal to a preset threshold precision; and determine masks based on the statistical value, wherein the merging is based on the weights in claim 14. But obtain a statistical value of first weights selected from among the weights based on having an inherent precision greater than or equal to a preset threshold precision does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 15 is dependent on claim 14 and includes all the limitations of claim 14. Claim 15 recites obtain a similarity to the statistical value for each of second weights, among the weights, having an inherent precision that is less than the preset threshold precision; and determine masks respectively corresponding to the second weights based on the similarity in claim 15. But obtain a similarity to the statistical value for each of second weights, among the weights, having an inherent precision that is less than the preset threshold precision does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 16 is dependent on claim 15 and includes all the limitations of claim 15. Claim 16 recites determine a binary mask that maximizes the similarity of each of the second weights in claim 16. But determine a binary mask that maximizes the similarity of each of the second weights does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 17 is dependent on claim 11 and includes all the limitations of claim 11. Claim 17 recites periodically train a dequantizer that performs the dequantizingin in claim 17. But periodically train a dequantizer that performs the dequantizingin weights does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim 18 recites the same limitations as claim 8 above. Therefore, claim
18 is rejected based on the same reasoning.
Claim 19 is dependent on claim 11 and includes all the limitations of claim 11. Claim 19 recites wherein the weights received from the clients are weights of neural network models individually trained by the clients in claim 17. But wherein the weights received from the clients are weights of neural network models individually trained by the client does not go beyond the abstract idea itself. There are no additional components in the claim that would make it significantly more than the abstract idea.
Claim Rejections 35 U.S.C. §103
6. 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 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.
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.
7. 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:
8. Claims 1-7, 9-17 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Khailany et al. (US 2022/0067512 A1) in view of Kim et al. (US Patent No. 12, 475, 192 B1).
Regarding claim 1 Khailany teaches a method, comprising:
receiving weights of respective clients, wherein each weight has a respectively corresponding precision, that is initially an inherent precision, (See Khailany paragraph [0050], A multiplier 122 multiplies the weight and input activation scale factors to produce a scale product factor s.sub.w(j)s.sub.a(j) that may be rounded to a desired precision by a rounding unit 124);
determining masks respectively corresponding to the weights based on the inherent precisions, (See Khailany paragraph [0030],The resulting increase in quantization error introduces significant accuracy loss for low-precision representations. The problem is exacerbated for DNNs whose input activations and/or weight values span a wide dynamic range),
based on the masks, determining an integrated weight by merging the weights having the reference precision, (See Khailany paragraph [0034], the floating-point weights and activations are collectively referred to as real values, and the quantized low-precision weights and activations collectively referred to as integer values); and
quantizing the integrated weight to generate quantized weights having the inherent precisions, respectively, (See Khailany paragraph [0034], the floating-point weights and activations are collectively referred to as real values, and the quantized low-precision weights and activations collectively referred to as integer values), and transmitting the quantized weights to the clients, (See Khailany paragraph [0166], provide data as input to a trained network, which can then generate one or more inferences as output…transmitted to client device 502).
Khailany does not explicitly disclose using a dequantizer to change the weights such that the precisions thereof are changed from the inherent precisions to a same reference precision.
However, Kim teaches using a dequantizer to change the weights such that the precisions thereof are changed from the inherent precisions to a same reference precision, (See Kim Col. 1 lines 64-67, Col. 2 lines 1-2, a method of dequantizing the quantized weight values and includes converting each of the quantized weight values expressed by a first number of bits to be expressed by a second number of bits and changing order of the weight values expressed by the second number of bits according to a dequantization pattern);
It would have been obvious to one with ordinary skill in the art before the
effective filing date of the claimed invention was made, to modify using a dequantizer to change the weights such that the precisions thereof are changed from the inherent precisions to a same reference precision of Kim for efficiently storing quantized weights of artificial intelligence models.
Regarding claim 2, Khailany taught the method of claim 1, as described above.
Khailany does not explicitly disclose wherein the dequantizer comprises: blocks, wherein each of the blocks has an input precision and an output precision.
However, Kim teaches wherein the dequantizer comprises: blocks, wherein each of the blocks has an input precision and an output precision, (See Kim Col. 7 lines 41-46, A dequantized 16-bit floating-point weight matrix 702 is matrix-multiplied by input values loaded to the register, and a 16-bit floating-point output value matrix 703 acquired as the result of the matrix multiplication is stored in the global memory).
It would have been obvious to one with ordinary skill in the art before the
effective filing date of the claimed invention was made, to modify wherein the dequantizer comprises: blocks, wherein each of the blocks has an input precision and an output precision of Kim for efficiently storing quantized weights of artificial intelligence models.
Regarding claim 3, Khailany taught the method of claim 2, as described above.
Khailany further teaches wherein the changing comprises:
inputting each of the weights to whichever of the blocks has an input precision that matches its inherent precision, (See Khailany paragraph [0032], The finer granularity at the vector level allows more precise scale factors to be determined based on local distribution of tensor parameter values in each vector of parameters 105…the unit (V) of a vector matches the unit of vector multiply-accumulate (MAC) hardware circuitry in a DNN accelerator); and
obtaining an output of whichever of the blocks has an output precision that matches the reference precision, (See Khailany paragraph [0058], the input activations are quantized, the computed output activation tensor is quantized to convert the higher-precision output activations back to N-bit vector elements with per-vector scale factors for the next layer).
Regarding claim 4, Khailany taught the method of claim 1, as described above.
Khailany further teaches wherein the determining of the masks comprises:
obtaining a statistical value of first weights, among the weights, (See Khailany paragraph [0038], For weights, scale factors may be determined using static calibration prior to inferencing. For activations, the scale factors may be determined using static calibration prior to inferencing or through dynamic calibration during inferencing), which have an inherent precision greater than or equal to a preset threshold precision among the weights;, (See Khailany paragraph [0066], the weight scale factors s.sub.w(j) are determined statically based on the trained model. Using static (max) calibration for weights and dynamic (max) calibration for activations has the potential to achieve significantly better accuracy using low bitwidths) and
determining the masks based on the statistical value, (See Khailany paragraph [0061], the distribution of values within the vector may lack enough samples to support the percentile and entropy calibration techniques to determine a statistically useful α).
Regarding claim 5, Khailany taught the method of claim 4, as described above.
Khailany further teaches wherein the determining of the masks based on the statistical value comprises, (See Khailany paragraph [0170], generating a trained model (e.g., trained deep neural network) that can be used to classify similar types of data, or generate other such inferences),
: for each of second weights of which an inherent precision is less than the statistical value among the weights, (See Khailany paragraph [0038], determines the a that minimizes the information loss between real and quantized distributions. For weights, scale factors may be determined using static calibration prior to inferencing g), obtaining a similarity thereof to the statistical value, (See Khailany paragraph [0170], a trained model (e.g., trained deep neural network) that can be used to classify similar types of data, or generate other such inferences); and
determining masks respectively corresponding to the second weights based on the similarities, (See Khailany paragraph [0170], generating a trained model (e.g., trained deep neural network) that can be used to classify similar types of data, or generate other such inferences).
Regarding claim 6 Khailany taught the method of claim 5, as described above.
Khailany further teaches wherein the determining of the masks respectively corresponding to the second weights comprises, (See Khailany paragraph [0037], ta scale factor for weights or activations is determined for every layer of the neural network. Known as per-layer scaling, a single scale factor is used for each weight tensor 103):
determining a binary mask that maximizes the similarities of the respective second weights, (See Khailany paragraph [0038], a calibration process is used to select the a used in Equation 1 for quantizing weights and activations. While a can be set to the maximum absolute value of the distribution (called max calibration)).
Regarding claim 7, Khailany taught the method of claim 1, as described above.
Khailany further teaches The method of claim 1, further comprising:
Khailany does not explicitly disclose training the dequantizer on a periodic basis. However, Kim teaches training the dequantizer on a periodic basis, (See Kim Col. 1 lines 60-62, performing matrix multiplication between a dequantized weight value matrix and the input value matrix).
It would have been obvious to one with ordinary skill in the art before the
effective filing date of the claimed invention was made, to modify training the dequantizer on a periodic basis of Kim for efficiently storing quantized weights of artificial intelligence models.
Regarding claim 9, Khailany taught the method of claim 1, as described above.
Khailany further teaches wherein the receiving of the weights comprises: receiving the weights of individually trained neural network models from the clients, (See Khailany paragraph [0093], a neural network model receives a first multi-dimensional input tensor of quantized parameters. In an embodiment, the quantized parameters are input activations or weights).
Regarding claim 10, Khailany taught the non-transitory computer-readable storage medium of claim, as described above. Khailany further teaches storing instructions that, when executed by a processor, cause the processor to perform the method of claim 1, (See Khailany paragraph [0052], a processor executing instructions stored in memory. The method 150 may also be embodied as computer-usable instructions stored on computer storage media).
Regarding claim 11, Khailany An electronic device comprising:
one or more processors, (See Khailany paragraph [0043], a processor);
a memory storing instructions configured to cause the one or more processors to, (See Khailany paragraph [0043], a processor executing instructions stored in memory):
receive weights of clients, wherein the weights have respectively corresponding precisions that are initially inherent precisions, (See Khailany paragraph [0050], A multiplier 122 multiplies the weight and input activation scale factors to produce a scale product factor s.sub.w(j)s.sub.a(j) that may be rounded to a desired precision by a rounding unit 124);
determine an integrated weight by merging the weights changed to have the reference precision, (See Khailany paragraph [0030],The resulting increase in quantization error introduces significant accuracy loss for low-precision representations. The problem is exacerbated for DNNs whose input activations and/or weight values span a wide dynamic range); and
quantize the integrated weight to weights respectively having the inherent precisions, (See Khailany paragraph [0034], the floating-point weights and activations are collectively referred to as real values, and the quantized low-precision weights and activations collectively referred to as integer values), and transmit the quantized weights to the clients, (See Khailany paragraph [0166], provide data as input to a trained network, which can then generate one or more inferences as output…transmitted to client device 502).
Khailany does not explicitly disclose dequantize the weights such that the precisions thereof are changed from the inherent precisions to a same reference precision.
However, Kim teaches dequantize the weights such that the precisions thereof are changed from the inherent precisions to a same reference precision, (See Kim Col. 1 lines 64-67, Col. 2 lines 1-2, a method of dequantizing the quantized weight values and includes converting each of the quantized weight values expressed by a first number of bits to be expressed by a second number of bits and changing order of the weight values expressed by the second number of bits according to a dequantization pattern);
It would have been obvious to one with ordinary skill in the art before the
effective filing date of the claimed invention was made, to modify dequantize the weights such that the precisions thereof are changed from the inherent precisions to a same reference precision of Kim for efficiently storing quantized weights of artificial intelligence models.
Regarding claim 12, Khailany taught the electronic device of claim 11, as described above.
Khailany does not explicitly disclose wherein the dequantizing is performed by a dequantizer comprising blocks, wherein each of the blocks has an input precision corresponding to at least one of the inherent precisions and has an output precision corresponding to at least one of the inherent precisions.
However, Kim teaches wherein the dequantizing is performed by a dequantizer comprising blocks, (See Kim Col. 1 lines 60-62, performing matrix multiplication between a dequantized weight value matrix and the input value matrix), wherein each of the blocks has an input precision corresponding to at least one of the inherent precisions and has an output precision corresponding to at least one of the inherent precisions, (See Kim Col. 7 lines 41-46, A dequantized 16-bit floating-point weight matrix 702 is matrix-multiplied by input values loaded to the register, and a 16-bit floating-point output value matrix 703 acquired as the result of the matrix multiplication is stored in the global memory).
It would have been obvious to one with ordinary skill in the art before the
effective filing date of the claimed invention was made, to modify wherein the dequantizing is performed by a dequantizer comprising blocks, wherein each of the blocks has an input precision corresponding to at least one of the inherent precisions and has an output precision corresponding to at least one of the inherent precisions of Kim for efficiently storing quantized weights of artificial intelligence models.
Regarding claim 13, Khailany taught the electronic device of claim 12, as described above. Khailany further teaches wherein the instructions are further configured to cause the one or more processors to, (See Khailany paragraph [0043], a processor executing instructions stored in memory):
input each of the weights to whichever of the blocks has an input precision corresponding to the weight's inherent precision; (See Khailany paragraph [0032], The finer granularity at the vector level allows more precise scale factors to be determined based on local distribution of tensor parameter values in each vector of parameters 105…the unit (V) of a vector matches the unit of vector multiply-accumulate (MAC) hardware circuitry in a DNN accelerator); and
obtain an output of whichever of the blocks has an output precision corresponding to the reference precision, (See Khailany paragraph [0058], the input activations are quantized, the computed output activation tensor is quantized to convert the higher-precision output activations back to N-bit vector elements with per-vector scale factors for the next layer).
Regarding claim 14, Khailany taught the electronic device of claim 11, as described above. Khailany further teaches wherein the instructions are further configured to cause the one or more processors to: (See Khailany paragraph [0043], a processor executing instructions stored in memory):
obtain a statistical value of first weights selected from among the weights, (See Khailany paragraph [0038], For weights, scale factors may be determined using static calibration prior to inferencing. For activations, the scale factors may be determined using static calibration prior to inferencing or through dynamic calibration during inferencing) based on having an inherent precision greater than or equal to a preset threshold precision; (See Khailany paragraph [0066], the weight scale factors s.sub.w(j) are determined statically based on the trained model. Using static (max) calibration for weights and dynamic (max) calibration for activations has the potential to achieve significantly better accuracy using low bitwidths) and
determine masks based on the statistical value, wherein the merging is based on the weights, (See Khailany paragraph [0061], the distribution of values within the vector may lack enough samples to support the percentile and entropy calibration techniques to determine a statistically useful α).
Regarding claim 15, Khailany taught the electronic device of claim 14, as described above. Khailany further teaches wherein the instructions are further configured to cause the one or more processors, to, (See Khailany paragraph [0043], a processor executing instructions stored in memory):
obtain a similarity to the statistical value for each of second weights, among the weights, (See Khailany paragraph [0038], determines the a that minimizes the information loss between real and quantized distributions. For weights, scale factors may be determined using static calibration prior to inferencing g), having an inherent precision that is less than the preset threshold precision, (See Khailany paragraph [0040], V should be selected to minimize the required number of scale factors...the precision of the vector-scaled approximation and resulting network accuracy); and
determine masks respectively corresponding to the second weights based on the similarity, (See Khailany paragraph [0170], generating a trained model (e.g., trained deep neural network) that can be used to classify similar types of data, or generate other such inferences).
Regarding claim 16, Khailany taught the electronic device of claim 15, as described above. Khailany further teaches wherein the instructions are further configured to cause the one or more processors to, (See Khailany paragraph [0043], a processor executing instructions stored in memory):
determine a binary mask that maximizes the similarity of each of the second weights, (See Khailany paragraph [0038], a calibration process is used to select the a used in Equation 1 for quantizing weights and activations. While a can be set to the maximum absolute value of the distribution (called max calibration)).
Regarding claim 17, Khailany taught the electronic device of claim 11, as described above. Khailany further teaches wherein the instructions are further configured to cause the one or more processors to, (See Khailany paragraph [0043], a processor executing instructions stored in memory).
Khailany does not explicitly disclose periodically train a dequantizer that performs the dequantizing.
However, Kim teaches periodically train a dequantizer that performs the dequantizing, (See Kim Col. 1 lines 58-62 , dequantizing the quantized weight values, copying an input value matrix to the register, and performing matrix multiplication between a dequantized weight value matrix and the input value matrix).
It would have been obvious to one with ordinary skill in the art before the
effective filing date of the claimed invention was made, to modify periodically train a dequantizer that performs the dequantizing of Kim for efficiently storing quantized weights of artificial intelligence models.
Regarding claim 19, Khailany taught the electronic device of claim 11, as described above. Khailany further teaches wherein the weights received from the clients are weights of neural network models individually trained by the clients, (See Khailany paragraph [0093], a neural network model receives a first multi-dimensional input tensor of quantized parameters…the quantized parameters are input activations or weights).
9. Claims 8 and 18 are rejected under 35 U.S.C. 103 as being unpatentable
over Khailany et al. (US 2022/0067512 A1) in view of Kim et al. (US Patent No. 12, 475, 192 B1) further in view of CRICRI et al. (US 2023/0164336 A1)
Regarding claim 8, Khailany taught the method of claim 7, as described above.
Khailany further teaches receiving learning weight data, (See Khailany paragraph [0026], The values of these filters are the weights that may be learned using a training dataset for a particular task);
generating pieces of quantized weight data by quantizing the learning weight data, (See Khailany paragraph [0026], the following description, the scaling and quantization techniques are described in the context of convolutional neural networks (CNNs) and this is not intended to be limiting. The convolutional layers represent the core of the CNN computation and are characterized by a set of filters that are usually 1×1 or 3×3, and occasionally 5×5 or larger. The values of these filters are the weights that may be learned using a training dataset for a particular task);
obtaining, for each of the blocks, a first loss that is determined based on a difference between intermediate output weight data predicted from a block and quantized weight data corresponding to the block, (See Khailany paragraph [0075], the vector sequencer 255 broadcasts a weight vector to each vector MAC unit 120 and sequences through multiple activation vectors before broadcasting another weight vector…produce intermediate values (e.g., scaled vector dot-product values)…When the output activations for a neural network layer have been computed and quantized by the quantization unit 210, the vector sequencer 255 may proceed to process a next layer by applying the output activations as input activations);
Khailany does not explicitly disclose wherein the dequantizer comprises: blocks,
wherein the training of the dequantizer comprises, obtaining a second loss that is determined based on a difference between final output weight data output from the dequantizer receiving the learning weight data, and true weight data corresponding to the learning weight data, and training the dequantizer based on the first loss and the second loss.
However, CRICR teaches wherein the dequantizer comprises: blocks,
wherein the training of the dequantizer comprises, (See CRICR paragraph [0090], dequantized M-features and the original video, the human-targeted decoder NN), obtaining a second loss that is determined based on a difference between final output weight data output from the dequantizer receiving the learning weight data, (See CRICRI paragraph [0097], determine a weight update for at least a subset of the coding pipeline based on the plurality of losses, wherein the weight update is configured to reduce a number of iterations for fine-tuning the coding pipeline for at least one of the plurality of tasks), and true weight data corresponding to the learning weight data, and training the dequantizer based on the first loss and the second loss, (See CRICRI paragraph [0093], The coding pipeline may further comprise other functions such as for example quantization and/or lossless encoding 1016 and lossless decoding and/or dequantization function 1026…The decoded features output by the decoder neural network 1024 may be input to one or more task neural networks).
It would have been obvious to one with ordinary skill in the art before the
effective filing date of the claimed invention was made, to modify wherein the dequantizer comprises: blocks, wherein the training of the dequantizer comprises, obtaining a second loss that is determined based on a difference between final output weight data output from the dequantizer receiving the learning weight data, and true weight data corresponding to the learning weight data, and training the dequantizer based on the first loss and the second loss of CRICRI to reduce a number of iterations for fine-tuning the coding pipeline for one of the task.
Regarding claim 18, Khailany taught the electronic device of claim 17, as described above. Khailany further teaches wherein the instructions are further configured to cause the one or more processors to: (See Khailany paragraph [0043], a processor executing instructions stored in memory),
receive learning weight data, (See Khailany paragraph [0026], The values of these filters are the weights that may be learned using a training dataset for a particular task);
generate pieces of quantized weight data by quantizing the learning weight data, (See Khailany paragraph [0026], the following description, the scaling and quantization techniques are described in the context of convolutional neural networks (CNNs) and this is not intended to be limiting. The convolutional layers represent the core of the CNN computation and are characterized by a set of filters that are usually 1×1 or 3×3, and occasionally 5×5 or larger. The values of these filters are the weights that may be learned using a training dataset for a particular task);
obtain a first loss that is determined based on a difference between intermediate output weight data predicted from a block and quantized weight data corresponding to the block, (See Khailany paragraph [0026], the following description, the scaling and quantization techniques are described in the context of convolutional neural networks (CNNs) and this is not intended to be limiting. The convolutional layers represent the core of the CNN computation and are characterized by a set of filters that are usually 1×1 or 3×3, and occasionally 5×5 or larger. The values of these filters are the weights that may be learned using a training dataset for a particular task);
Khailany together with Kim does not explicitly disclose wherein the dequantizer comprises: blocks, for each of the plurality of blocks; and obtain a second loss that is determined based on a difference between final output weight data output, from the dequantizer receiving the learning weight data and true weight data corresponding to the learning weight data; and train the dequantizer based on the first loss and the second loss.
However, CRICR teaches wherein the dequantizer comprises: blocks, (See CRICR paragraph [0090], dequantized M-features and the original video, the human-targeted decoder NN): for each of the plurality of blocks; and obtain a second loss that is determined based on a difference between final output weight data output, (See CRICRI paragraph [0097], determine a weight update for at least a subset of the coding pipeline based on the plurality of losses, wherein the weight update is configured to reduce a number of iterations for fine-tuning the coding pipeline for at least one of the plurality of tasks), from the dequantizer receiving the learning weight data and true weight data corresponding to the learning weight data; and train the dequantizer based on the first loss and the second loss, (See CRICRI paragraph [0093], The coding pipeline may further comprise other functions such as for example quantization and/or lossless encoding 1016 and lossless decoding and/or dequantization function 1026…The decoded features output by the decoder neural network 1024 may be input to one or more task neural networks).
It would have been obvious to one with ordinary skill in the art before the
effective filing date of the claimed invention was made, to modify wherein the dequantizer comprises: blocks, for each of the plurality of blocks; and obtain a second loss that is determined based on a difference between final output weight data output, from the dequantizer receiving the learning weight data and true weight data corresponding to the learning weight data; and train the dequantizer based on the first loss and the second loss of CRICRI to reduce a number of iterations for fine-tuning the coding pipeline for one of the task.
Conclusions/Points of Contacts
The prior art made of record and not relied upon is considered pertinent to
applicant’s disclosure. See form PTO-892.
Sriram et al. (US 2022/0044114 A1), an object detection or classification network is trained using full floating-point precision (either single-precision 32-bit floating point or double-precision 64-bit floating point) to represent the weights and activations of a deep neural network (DNN). However, when these trained models are deployed to edge devices that have less computational resources.
Han et all. (US 20230143985 A1) provides a data feature extraction method and a related apparatus, to reduce computing resources required for performing feature extraction on input information based on a quantized neural network. To be specific, a storage resource for storing a related parameter of the neural network is saved, and further, a computing resource required for performing feature extraction based on the related parameter of the neural network can be saved, thereby reducing a limitation on an application of neural network-based artificial intelligence to a resource-limited device. Any inquiry concerning this communication or earlier communications from the examiner should be directed to MULUEMEBET GURMU whose telephone number is (571)270-7095. The examiner can normally be reached M-F 9am - 5pm.
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, Tony Mahmoudi can be reached at 5712724078. 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.
/MULUEMEBET GURMU/Primary Examiner, Art Unit 2163