Detailed Action
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 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.
Claim(s) 1 – 2 and 12 – 13 rejected under 35 U.S.C. 103 as being unpatentable over Park et al., "Dual-Precision Deep Neural Network" in view of CHOUKROUN et al., Pub. No.: WO2020160787A1.
Regarding claim 1, Park teaches: access a first machine learning model comprising a first plurality of blocks, the first plurality of blocks being associated with a first precision and comprising a first block; access a second machine learning model
(Park, (page: 31, section 3) “The proposed method creates dual-precision DNNs by training both the low- precision (b bits) and high-precision (b+1 bits) parts of the weights. We propose weight-bit sharing between the two precision modes, where the b-bit values of the weights in the low precision mode [access a first machine learning model comprising a first plurality of blocks, the first plurality of blocks being associated with a first precision and comprising a first block] are directly used for the b high-order bits of the weights in the high-precision mode (Fig 2(a)) [access a second machine learning model]. This bit sharing allows efficient transition between the two models during inference, realized with simple appending and removing of the 1 up-scaling bits.”)
… comprising a second plurality of blocks associated with a second precision different from the first precision
(Park, (page: 31), “3.1 Quantization Method Quantization of DNN weights is a process of mapping the weights into different levels, I R^3 based on the number of levels and the scale s R^1 [… comprising a second plurality of blocks associated with a second precision different from the first precision], which are determined by the target precision. Generally, if two DNN models are individually trained for two different target precision, there will be no dependencies between the weight values of the two models.”)
wherein: the second plurality of blocks comprises a first block; and the first block of the second plurality of blocks corresponds to the first block of the first plurality of blocks;
(Park, (page: 31, section 3.1) “To integrate two precision modes in a single set of weights, the training method should maintain dependencies between the values in the weights with the two precision modes. To address this, the proposed training method builds the up-scaled b+1-bit model based on the original b-bit model [wherein: the second plurality of blocks comprises a first block; and the first block of the second plurality of blocks corresponds to the first block of the first plurality of blocks] (i.e.: a higher precision model derived from the low precision model. The higher precision model retains the same network architecture as the lower precision model, such that each layer or block of the higher precision corresponds to the lower precision model), while maintaining the weight levels and values of the b shared bits.”)
process an input to the quantized machine learning model using the first plurality of blocks of the first machine learning model and the second plurality of blocks of the second machine learning model,
(Park, (page: 30), “In this paper, we propose a new method to train dual-precision DNNs that can switch between the two precision modes without the need for re-training. By sharing the common bits (b bits) [process an input] between the low - and high-precision weights, the dual-precision DNN allows dynamic on-line switch between the two precision modes [to the quantized machine learning model using the first plurality of blocks of the first machine learning model and the second plurality of blocks of the second machine learning model,] only with truncation (precision down-scaling) or concatenation (precision up-scaling) of the last 1 bits.”)
Park does not teach:
one or more memories comprising computer-executable instructions; and one or more processors configured to execute the computer-executable instructions and cause the processing system to
wherein, to process the input, the one or more processors are configured to execute the computer-executable instructions and cause the processing system to modify an output of the first block of the first plurality of blocks based on the corresponding first block of the second plurality of blocks; and provide an output of the first machine learning model based on the processing.
CHOUKROUN teaches:
one or more memories comprising computer-executable instructions; and one or more processors configured to execute the computer-executable instructions and cause the processing system to:
(CHOUKROUN, (page: 11, line [12 – 15]), “The inputs received via the I/O interface are then processed by a code stored in the memory storage 106, by execution of the code by the one or more processor(s) 108 [A processing system comprising: one or more memories comprising computer-executable instructions; and one or more processors configured to execute the computer-executable instructions and cause the processing system to]. The code contains instructions for a process for configuring a NN, either based on NN weights or based on NN activations.”)
wherein, to process the input, the one or more processors are configured to execute the computer-executable instructions and cause the processing system to modify an output of the first block of the first plurality of blocks based on the corresponding first block of the second plurality of blocks; and
(CHOUKROUN, (page: 2, line [7 – 15]), “According to a second aspect of the invention, a system for configuring a neural network, trained from a plurality of data samples, comprising: processing circuitry, configured to: quantizing each layer of the neural network to produce a quantized neural network with a plurality of respective scaling factors; locating one or more layers of the quantized neural network; computing a modified quantization for the one or more located layers to produce a modified quantized neural network [wherein, to process the input, the one or more processors are configured to execute the computer-executable instructions and cause the processing system to modify an output of the first block of the first plurality of blocks]; and adjusting the plurality of scaling factors of the modified quantized neural network by computing a similarity between a plurality of neural network outputs and a plurality of modified quantized neural network outputs [based on the corresponding first block of the second plurality of blocks].”)
provide an output of the first machine learning model based on the processing.
(CHOUKROUN, (page: 11, line [12 – 17]), “The inputs received via the I/O interface are then processed by a code stored in the memory storage 106, by execution of the code by the one or more processor(s) 108. The code contains instructions for a process for configuring a NN, either based on NN weights or based on NN activations. Outcomes of NN configuration are outputted via the I/O interface 104 [provide an output of the first machine learning model based on the processing] by the one or more processor(s) executing the code instructions, whereas the outputs may be directed to back to the client(s)”)
CHOUKROUN and Park are related to the same field of endeavor (i.e.: machine learning optimization). It would have been obvious to a person having ordinary skill in the art before the effective filing date of the invention to combine the teaching of CHOUKROUN with teachings of Park to add a selective quantization optimization process that modifies the quantization of specific layers and adjusts their scaling factors based on the similarity between the outputs of the original and quantized neural networks to improve the accuracy of the quantized model (CHOUKROUN, Abstract).
Claim 12, recites limitations analogous to claim 1, so is rejected under the same rationale.
Regarding claim 2, Park in view of CHOUKROUN teach the method of claim 1.
Park further teaches: wherein the second precision is higher than the first precision.
(Park, (page: 31), “3. DUAL-PRECISION DEEP NEURAL NETWORK The proposed method creates dual-precision DNNs by training both the low- precision (b bits) and high-precision (b+1 bits) parts of the weights [wherein the second precision is higher than the first precision]. We propose weight-bit sharing between the two precision modes, where the b-bit values of the weights in the low precision mode are directly used for the b high-order bits of the weights in the high-precision mode (Fig 2(a)).”)
Claim 13, recites limitations analogous to claim 2, so is rejected under the same rationale.
Claim(s) 3 and 14 rejected under 35 U.S.C. 103 as being unpatentable over Park in view of CHOUKROUN and in further view of Cohen et al., Pub. No.: US20190102671A1.
Regarding claim 3, Park in view of CHOUKROUN teach the method of claim 2.
Park in view of CHOUKROUN do not teach:
wherein the second precision corresponds to a 16-bit bit width and wherein the first precision corresponds to a 4-bit bit width.
Cohen teaches:
wherein the second precision corresponds to a 16-bit bit width and
(Cohen, “[0296] There is further disclosed a CNN accelerator, wherein the high-precision CNN is selected from 8-bit integer and 16-bit floating point [wherein the second precision corresponds to a 16-bit bit width].”)
wherein the first precision corresponds to a 4-bit bit width.
(Cohen, “[0025] A CNN generally includes several layers in each topology. These layers include convolution layers, pooling layers, and reduction layers. While convolution and pooling layers are well defined for any data type, the reduction layer can be improved by defining it for topologies with low precision data types, such as 1-bit, 2-bit, or 4-bit [wherein the first precision corresponds to a 4-bit bit width]. ”)
Cohen, Park and CHOUKROUN are related to the same field of endeavor (i.e.: machine learning optimization). It would have been obvious to a person having ordinary skill in the art before the effective filing date of the invention to combine the teaching of Cohen with teachings of Park and CHOUKROUN to add CNN accelerator architecture to performs multi-layer convolution to accelerate inference (Cohen, Abstract).
Claim 14, recites limitations analogous to claim 3, so is rejected under the same rationale.
Claim(s) 4 and 15 rejected under 35 U.S.C. 103 as being unpatentable over Park in view of CHOUKROUN and in further view of Golub et al., Pub. No.: US20220245444A1.
Regarding claim 4, Park in view of CHOUKROUN teach the method of claim 1.
Park in view of CHOUKROUN do not teach:
wherein: the first machine learning model has a first size; the second machine learning model has a second size; and the second size is smaller than the first size.
Golub teaches:
wherein: the first machine learning model has a first size
(Golub, “[0054] At 402, one or more control processor(s) may receive a first neural network model having a first model size (e.g., a smaller size) for a given application [wherein: the first machine learning model has a first size;]. At 404, control processor(s) may configure the first model, based on training parameter(s), to execute first training process(es) on one or more AI processor(s). At 406, control processor(s) may monitor a plurality of statistics at various locations of the first model produced upon execution of the first training process(es) by the AI processor(s).”)
the second machine learning model has a second size; and the second size is smaller than the first size.
(Golub, “[0056] Returning to the example of FIG. 4, at 410, control processor(s) may receive the second neural network model having a second model size (e.g., a larger size) [the second machine learning model has a second size; and the second size is smaller than the first size]. At 412, control processor(s) may load the mappings generated at 408 to configure the second model to execute second training process(es) on AI processor(s) using training parameters adjusted according to the mappings.”)
Golub, Park and CHOUKROUN are related to the same field of endeavor (i.e.: machine learning optimization). It would have been obvious to a person having ordinary skill in the art before the effective filing date of the invention to combine the teaching of Golub with teachings of Park and CHOUKROUN to add adaptive training parameter optimization based on monitoring training statistics and dynamically adjusting training parameters to improve training stability and maintain desired training performance (Golub, Abstract).
Claim 15, recites limitations analogous to claim 4, so is rejected under the same rationale.
Claim(s) 5 and 16 rejected under 35 U.S.C. 103 as being unpatentable over Park in view of CHOUKROUN and in further view of Kwon et al. "Heterogeneous dataflow accelerators for multi-dnn workloads."
Regarding claim 5, Park in view of CHOUKROUN teach the method of claim 1.
Park in view of CHOUKROUN do not teach:
wherein: the first machine learning model is accessed by a first circuit of an integrated-circuit (IC) device; and the second machine learning model is accessed by a second circuit of the IC device different from the first circuit.
Kwon teaches:
wherein: the first machine learning model is accessed by a first circuit of an integrated-circuit (IC) device; and the second machine learning model is accessed by a second circuit of the IC device different from the first circuit.
(Kwon, (page: 2), “In this work, we propose a new class of DNN accelerators called heterogeneous dataflow accelerators (HDAs). HDAs provide flexibility by employing multiple sub-accelerators [wherein: the first machine learning model… … and the second machine learning model], each tuned for a different dataflow, within an accelerator chip [… is accessed by a first circuit of an integrated-circuit (IC) device… … is accessed by a second circuit of the IC device different from the first circuit]. HDAs provide two important features: (i) dataflow flexibility, enabled by scheduling each layer from the multiple DNN models on the most efficient sub-accelerator for each layer, as long as possible. (ii) high utilization, enabled by scheduling multiple layers from different models across the sub-accelerators simultaneously.”)
Kwon, Park and CHOUKROUN are related to the same field of endeavor (i.e.: machine learning optimization). It would have been obvious to a person having ordinary skill in the art before the effective filing date of the invention to combine the teaching of Kwon with teachings of Park and CHOUKROUN to enhance hardware resource optimization and workload scheduling across multiple neural network accelerators to improve execution efficiency for multi-model workloads. (Kwon, Abstract).
Claim 16, recites limitations analogous to claim 5, so is rejected under the same rationale.
Claim(s) 6 and 17 rejected under 35 U.S.C. 103 as being unpatentable over Park in view of CHOUKROUN and in further view of Chen et al. "Data-free quantization via mixed-precision compensation without fine-tuning."
Regarding claim 6, Park in view of CHOUKROUN teach the method of claim 1.
Park in view of CHOUKROUN do not teach:
wherein: the first machine learning model was generated by quantizing a baseline machine learning model having a baseline precision higher than the first precision, and the second machine learning model was trained to adjust for quantization errors resulting from the quantization of the baseline machine learning model.
Chen teaches:
wherein: the first machine learning model was generated by quantizing a baseline machine learning model having a baseline precision higher than the first precision, and the second machine learning model was trained to adjust for quantization errors resulting from the quantization of the baseline machine learning model.
(Chen, page: 2, “In two adjacent layers of a neural network, we assume that the quantized error caused by a low-precision quantized layer can be restored via the reconstruction of a high-precision quantized layer [the second machine learning model was trained to adjust for quantization errors resulting from the quantization of the baseline machine learning model]. Specifically, we quantize the weights in one layer into low precision values (e.g., 2-bit) and then recover the performance by reconstructing relatively higher precision (e.g., 6-bit) weights in the next layer [the first machine learning model was generated by quantizing a baseline machine learning model having a baseline precision higher than the first precision] (i.e.: a baseline higher precision model (weight) being quantized to produce lower precision by reducing bit precision during quantization). The layer-wise mixed-precision compensation assumption is described in Section 4.1”)
Chen, Park and CHOUKROUN are related to the same field of endeavor (i.e.: machine learning optimization). It would have been obvious to a person having ordinary skill in the art before the effective filing date of the invention to combine the teaching of Chen with teachings of Park and CHOUKROUN to add a mixed precision compensation technique to improve the accuracy of low precision quantized neural networks without requiring training data or additional fine tuning (Chen, Abstract).
Claim 17, recites limitations analogous to claim 6, so is rejected under the same rationale.
Claim(s) 7 and 18 rejected under 35 U.S.C. 103 as being unpatentable over Park in view of CHOUKROUN and in further view of Ryoo et al., Pub. No.: US20230114556A1.
Regarding claim 7, Park in view of CHOUKROUN teach the method of claim 1.
Park in view of CHOUKROUN do not teach:
wherein: the first plurality of blocks comprises an ordered network of blocks, the first plurality of blocks further comprises a second block configured to receive as an input the modified output of the first block of the first plurality of blocks and to process the received input, and to process the input, the one or more processors are configured to further execute the computer-executable instructions and cause the processing system to modify an output of the second block of the first plurality of blocks using a corresponding second block of the second plurality of blocks.
Ryoo teaches:
wherein: the first plurality of blocks comprises an ordered network of blocks,
(Ryoo, “[0036] The neural network system 100 processes a network input 102 using one or more blocks arranged in levels to generate a network output 104 that characterizes the network input. The one or more blocks are arranged in an ordered sequence of levels such that each block belongs to only one of the levels [the first plurality of blocks comprises an ordered network of blocks]. Each block of the one or more blocks is configured to process a block input using one or more neural network layers to generate a block output.”)
the first plurality of blocks further comprises a second block configured to receive as an input the modified output of the first block of the first plurality of blocks and to process the received input,
(Ryoo, “[0015] In some implementations, the target block is associated with a target level, and the target block receives: (i) a respective first block output of each first block that is associated with a level that precedes the target level, and (ii) a respective second block output of each second block that is associated with a level that precedes the target level [the first plurality of blocks further comprises a second block configured to receive as an input the modified output of the first block of the first plurality of blocks and to process the received input].”)
and to process the input, the one or more processors are configured to further execute the computer-executable instructions and cause the processing system to modify an output of the second block of the first plurality of blocks using a corresponding second block of the second plurality of blocks.
(Ryoo, “[0079] The system scales each second block output by a function of a corresponding attention weight (404) [to process the input, the one or more processors are configured to further execute the computer-executable instructions and cause the processing system to modify an output of the second block of the first plurality of blocks using a corresponding second block of the second plurality of blocks] (i.e.: the second block outputs are modified by scaling each output according to a corresponding attention weight). The corresponding attention weights are learnable parameters which can be trained, e.g., by training engine 108 of FIG. 1 , and each attention weight corresponds to a second block output. In one example, the system can apply a softmax function to the attention weights corresponding to each second block output, then scale each second block output by the corresponding attention weight output by the softmax function. Using a softmax function can emphasize the contribution of the most impactful second block or blocks.”)
Ryoo, Park and CHOUKROUN are related to the same field of endeavor (i.e.: machine learning optimization). It would have been obvious to a person having ordinary skill in the art before the effective filing date of the invention to combine the teaching of Ryoo with teachings of Park and CHOUKROUN to add a mechanism that adaptively weights and combines intermediate outputs from multiple neural network blocks before processing by a target block. (Ryoo, Abstract).
Claim 18, recites limitations analogous to claim 7, so is rejected under the same rationale.
Claim(s) 8, 11, 19 and 22 rejected under 35 U.S.C. 103 as being unpatentable over Park in view of CHOUKROUN and in further view of Kim et al., "Domain adaptation without source data."
Regarding claim 8, Park in view of CHOUKROUN teach the method of claim 1.
Park in view of CHOUKROUN do not teach:
wherein: the first machine learning model was trained based on training data from a source domain, the second machine learning model was trained using adjustment data from the source domain, and the second machine learning model was trained without using labels for the adjustment data.
Kim teaches:
wherein: the first machine learning model was trained based on training data from a source domain,
(Kim, (page: 2), “To address the issue of reliable, yet, very few target samples, we propose a new framework consisting of two parts. One is a pre-trained model from the source domain [the first machine learning model was trained based on training data from a source domain] where all the weights are frozen, and the other is a target model that is initialized from the pre-trained source model but evolves progressively by optimizing two losses (see Fig. 2).”)
the second machine learning model was trained using adjustment data from the source domain, and the second machine learning model was trained without using labels for the adjustment data.
(Kim, (page: 9), “In this paper, we have proposed a novel paradigm shift for unsupervised domain adaption, called Source data-Free Domain Adaptation (SFDA) from a source pre-trained model. Our main (target) model utilizes a pre-trained model from the source domain instead of using source data directly. Specifically, two types of pseudo labels are used for training our target model [the second machine learning model was trained using adjustment data from the source domain]. Target-oriented pseudo labels obtained from the adaptive prototype memory are used to train the target model [the second machine learning model was trained without using labels for the adjustment data] in a self-learning manner while source-oriented pseudo labels prevent the target model from a self-biasing problem.”)
Kim, Park and CHOUKROUN are related to the same field of endeavor (i.e.: machine learning optimization). It would have been obvious to a person having ordinary skill in the art before the effective filing date of the invention to combine the teaching of Kim with teachings of Park and CHOUKROUN to add source domain adaption technique that adapts a pre trained model to a target domain using pseudo labels and target domain samples without requiring to access to the original training data to enhance model generalization to new domains while preventing data privacy and eliminating the need for source domain data during adaption. (Kim, Abstract).
Claim 19, recites limitations analogous to claim 8, so is rejected under the same rationale.
Regarding claim 11, Park in view of CHOUKROUN and Kim teach the method of claim 8.
Kim further teaches: wherein: the one or more processors are configured to further execute the computer-executable instructions and cause the processing system to adapt the second machine learning model to a target domain based on labeled adaptation data for the target domain, and
(Kim, page: 3, “Once the pseudo labels ˆyt are obtained through adaptive prototype memory, we additionally remove unreliable samples through set-to-set distance-based confidence. Overall, we adapt the target model to the target domain using pseudo labels ˆyt in a self-learning manner [second machine learning model to a target domain based on labeled adaptation data for the target domain], while using source knowledge ˆ ys from the pre-trained source model as a regularizer.”)
the first machine learning model is frozen during adaptation to the target domain.
(Kim, page: 2, “To address the issue of reliable, yet, very few target samples, we propose a new framework consisting of two parts. One is a pre-trained model from the source domain where all the weights are frozen [the first machine learning model is frozen during adaptation to the target domain], and the other is a target model that is initialized from the pre-trained source model but evolves progressively by optimizing two losses (see Fig. 2).”)
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 Kim with teachings of Park and CHOUKROUN for the same reasons disclosed for claim 8.
Claim 22, recites limitations analogous to claim 11, so is rejected under the same rationale.
Claim(s) 23 and 29 – 30 rejected under 35 U.S.C. 103 as being unpatentable over Park in view of Nakata et al., Pub. No.: US20230196142A1 and Chen.
Regarding claim 23, Park teaches: A processor-implemented method, comprising: accessing a first machine learning model comprising a first plurality of blocks; generating a second machine learning model
(Park, (page: 31, section 3) “The proposed method creates dual-precision DNNs by training both the low- precision (b bits) and high-precision (b+1 bits) parts of the weights. We propose weight-bit sharing between the two precision modes, where the b-bit values of the weights in the low precision mode [accessing a first machine learning model comprising a first plurality of blocks] are directly used for the b high-order bits of the weights in the high-precision mode (Fig 2(a)) [generating a second machine learning model]. This bit sharing allows efficient transition between the two models during inference, realized with simple appending and removing of the 1 up-scaling bits.”)
comprising a second plurality of blocks by quantizing the first machine learning model;
(Park, (page: 31), “3.1 Quantization Method Quantization of DNN weights [by quantizing the first machine learning model] is a process of mapping the weights into different levels, I R^3 based on the number of levels and the scale s R^1 [comprising a second plurality of blocks], which are determined by the target precision. Generally, if two DNN models are individually trained for two different target precision, there will be no dependencies between the weight values of the two models.”)
Park does not teach:
training a third machine learning model comprising a third plurality of blocks for adjusting for the quantization of the first machine learning model; and
deploying the second machine learning model and the third machine learning model for inferencing.
Nakata teaches:
training a third machine learning model comprising a third plurality of blocks for adjusting for the quantization of the first machine learning model; and
(Nakata, “[0035] With this, the computing of the second inference model is started on the basis of the settings information for quantization and the first inference model. This thus enables to find, at an early stage, the third inference model that is based on quantization and the first inference model [training a third machine learning model comprising a third plurality of blocks for adjusting for the quantization of the first machine learning model].”)
Nakata and Park are related to the same field of endeavor (i.e.: machine learning optimization). It would have been obvious to a person having ordinary skill in the art before the effective filing date of the invention to combine the teaching of Nakata with teachings of Park to add model scaling, quantization and performance based validation to source free domain adaption framework to improve model capacity. (Nakata, Abstract).
Park in view of Nakata do not teach:
deploying the second machine learning model and the third machine learning model for inferencing.
Chen teaches:
deploying the second machine learning model and the third machine learning model for inferencing
(Chen, (col. 5 line [47 – 56]), “In this scenario, the user 118 may cause the client device 120 to send, at circle (2), a request to deploy the ML model(s) 108 to one or more edge devices 122A-122N [deploying the second machine learning model and the third machine learning model for inferencing]. The request may identify a storage location where the model is located (e.g., a URL/URI where the model file or files are available, which may be within or outside of the provider network 100), as well as identify a particular edge device 122A, multiple individual edge devices 122A-122N, or a group of edge devices (e.g., “store_security_cameras”) to deploy the ML model 108 to.”)
Chen, Park and Nakata are related to the same field of endeavor (i.e.: machine learning optimization). It would have been obvious to a person having ordinary skill in the art before the effective filing date of the invention to combine the teaching of Chen with teachings of Park and Nakata to add a mixed precision compensation technique to improve the accuracy of low precision quantized neural networks without requiring training data or additional fine tuning (Chen, Abstract).
Regarding claim 29, Park in view of Nakata and Chen teach the method of claim 23.
Park further teaches: wherein: parameters of the second machine learning model are encoded using a first value representation,
(Park, (page: 32), “If the model does not utilize enough levels, the accuracy of the model cannot be high enough. Therefore, to facilitate level branching, we apply index normalization. More specifically, the index parameters are initialized to a normal distribution with a mean of 0 and a standard deviation of 0.3 (m = 0, σ = 0.3), and the range of the updated parameters is scaled as same as the range of the index parameters [parameters of the second machine learning model are encoded using a first value representation].”)
the second value representation has a higher precision than the first value representation.
(Park, (page: 32), “The final hypothesis is then compared with the label corresponding to the training data to calculate the gradient updates. Weight update is performed differently in odd and even epochs. In odd epochs, only the shared b-bit low- precision parameters are updated. In even epochs, on the other hand, 1-bit level parameters λ added in the high-precision weights [the second value representation has a higher precision than the first value representation] are updated as well as the shared b-bit parameters.”)
Nakata further teaches: parameters of the third machine learning model are encoded using a second value representation, and
(Nakata, “[0035] With this, the computing of the second inference model is started on the basis of the settings information for quantization and the first inference model. This thus enables to find, at an early stage, the third inference model that is based on quantization and the first inference model [parameters of the third machine learning model are encoded using a second value representation].”)
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 Nakata with teachings of Park and Chen for the same reasons disclosed for claim 23.
Regarding claim 30, Park in view of Nakata and Chen teach the method of claim 23.
Chen further teaches: wherein deploying the second machine learning model and the third machine learning model for inferencing comprises: deploying the second machine learning model to be executed on a first hardware component and deploying the third machine learning model to be executed on a second hardware component.
(Chen, (col. 8 line [17 – 24]), “ Notably, in some embodiments the model optimizer 114B and/or inference engine 132 may utilize a common set of APIs (across disparate types of deployments) to optimize a model, load a model, and/or perform inference, allowing the application 127 to be easily written to interact with the model optimizer 114B and/or inference engine 132, and flexibly be deployed in a number of different hardware environments [deploying the second machine learning model to be executed on a first hardware component and deploying the third machine learning model to be executed on a second hardware component].”)
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 Chen with teachings of Park and Nakata for the same reasons disclosed for claim 23.
Claim(s) 24 – 25 and 28 rejected under 35 U.S.C. 103 as being unpatentable over Park in view of Nakata, Chen and in further view of Kim.
Regarding claim 24, Park in view of Nakata, Chen teach the method of claim 23.
Park in view of Nakata, Chen do not teach:
wherein each of the first plurality of blocks comprises at least one of a layer of the first machine learning model or a transformer of the first machine learning model.
Kim teaches:
wherein each of the first plurality of blocks comprises at least one of a layer of the first machine learning model or a transformer of the first machine learning model.
((Kim, page: 33), “Fig 4(b) represents the maximum number of index levels in all the layers of the same DNN model [wherein each of the first plurality of blocks comprises at least one of a layer of the first machine learning model] as in Fig 4(a) during the proposed training process. It shows that level branching according to the added 1-bit occurs right after the transition to the phase 2. This is because the phase used η to more focus on the shared 2-bit weights, creating little influence on index parameters to be updated.”)
Kim, Park, Nakata and Chen are related to the same field of endeavor (i.e.: machine learning optimization). It would have been obvious to a person having ordinary skill in the art before the effective filing date of the invention to combine the teaching of Kim with teachings of Park, Nakata and Chen to add source domain adaption technique that adapts a pre trained model to a target domain using pseudo labels and target domain samples without requiring to access to the original training data to enhance model generalization to new domains while preventing data privacy and eliminating the need for source domain data during adaption. (Kim, Abstract).
Regarding claim 25, Park in view of Nakata, Chen teach the method of claim 23.
Nakata further teaches: the third machine learning model is trained using adjustment data from the source domain, and
(Nakata, “[0035] With this, the computing of the second inference model is started on the basis of the settings information for quantization and the first inference model. This thus enables to find, at an early stage, the third inference model that is based on quantization and the first inference model [the third machine learning model is trained using adjustment data from the source domain].”)
the third machine learning model is trained without using labels for the adjustment data.
(Nakata, “[0240] Next, processor 301 trains the third inference model, using machine learning (S704). Processor 301 then determines whether the performance of the trained third inference model satisfies a condition (S705) [the third machine learning model is trained without using labels for the adjustment data]. Subsequently, processor 301 outputs the trained third inference model when the performance satisfies the condition (S706).”)
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 Nakata with teachings of Park and Chen for the same reasons disclosed for claim 23.
Park in view of Nakata, Chen do not teach:
wherein: the first machine learning model was trained based on training data from a source domain, the third machine learning model is trained using adjustment data from the source domain, and the third machine learning model is trained without using labels for the adjustment data.
Kim teaches:
wherein: the first machine learning model was trained based on training data from a source domain,
(Kim, (page: 2), “To address the issue of reliable, yet, very few target samples, we propose a new framework consisting of two parts. One is a pre-trained model from the source domain [wherein: the first machine learning model was trained based on training data from a source domain] where all the weights are frozen, and the other is a target model that is initialized from the pre-trained source model but evolves progressively by optimizing two losses (see Fig. 2).”)
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 Kim with teachings of Park, Nakata and Chen for the same reasons disclosed for claim 24.
Regarding claim 28, Park in view of Nakata, Chen and Kim teach the method of claim 25.
Nakata further teaches: further comprising adapting the third machine learning model to a target domain based on labeled adaptation data for the target domain,
(Nakata, “[0240] Next, processor 301 trains the third inference model, using machine learning (S704). Processor 301 then determines whether the performance of the trained third inference model satisfies a condition (S705). Subsequently, processor 301 outputs the trained third inference model when the performance satisfies the condition (S706) [adapting the third machine learning model to a target domain based on labeled adaptation data for the target domain].”)
wherein the second machine learning model is frozen during adaptation to the target domain.
(Kim, (page: 2), “To address the issue of reliable, yet, very few target samples, we propose a new framework consisting of two parts. One is a pre-trained model from the source domain where all the weights are frozen [wherein the second machine learning model is frozen during adaptation to the target domain], and the other is a target model that is initialized from the pre-trained source model but evolves progressively by optimizing two losses (see Fig. 2).”)
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 Nakata with teachings of Park, Kim and Chen for the same reasons disclosed for claim 24.
Allowable subject matter
Claim(s) 9, 10, 20, 21, 26 and 27 objected to as being dependent upon a rejected base claim and would be allowable if rewritten in independent form including all of the limitations of the base claim. The prior art made of record does not teach, make obvious, or suggest the claim limitations as disclosed in applicant's claims.
Claim 9 and analogues claim 20 recite:
wherein, to train the second machine learning model, the one or more processors are configured to execute the computer-executable instructions and cause the processing system to generate an adjustment loss for a first block of the second machine learning model based on: (i) a first feature map generated by a first block of a baseline machine learning model based on a first exemplar in the adjustment data; (ii) a second feature map generated by a quantized version of the first block of the baseline machine learning model based on the first exemplar, the quantized version corresponding to the first block of the first plurality of blocks; and (iii) a third feature map generated by the first block of the second plurality of blocks based on the first exemplar.
Closest prior art(s):
Park et al., "Dual-Precision Deep Neural Network".
Park teaches train dual-precision DNNs that can switch between the two precision modes without the need for re-training. By sharing the common bits (b bits) between the low- and high-precision weights, the dual-precision DNN allows dynamic on-line switch between the two precision modes only with truncation (precision down-scaling) or concatenation (precision up-scaling) of the last 1 bits. However, Park does not teach a second machine learning model is trained by generating an adjustment loss for its first block using adjustment data. The adjustment loss compares three feature maps produced from the same exemplar: a feature map from the corresponding block of a baseline model, a feature map from a quantized version of that baseline block, and a feature map from the corresponding block of the second machine learning model. This loss guides the second model to learn adjustments that compensate for quantization effects.
Kim et al., "Domain adaptation without source data."
Kim discusses source data free domain adaptation (SFDA), a domain adaption approach that eliminates the need to access source domain training data. Instead, it uses a pre trained source model and progressively updates a target model through self-learning. The method leverages low self-entropy target samples to correctly classified, to facilitate reliable adaption while preserving source domain privacy. However, Kim does not teach a second machine learning model is trained by generating an adjustment loss for its first block using adjustment data. The adjustment loss compares three feature maps produced from the same exemplar: a feature map from the corresponding block of a baseline model, a feature map from a quantized version of that baseline block, and a feature map from the corresponding block of the second machine learning model. This loss guides the second model to learn adjustments that compensate for quantization effects.
Claim 10 and analogues claim 21 recite:
wherein to train the second machine learning model, the one or more processors are configured to execute the computer-executable instructions and cause the processing system to generate an adjustment loss for a first block of the second plurality of blocks based on: (i) a first model output generated by a baseline machine learning model based on a first exemplar in the adjustment data; (ii) a second model output generated by the first machine learning model based on the first exemplar; and (iii) a third model output generated by the second machine learning model based on the first exemplar.
Closest prior art(s):
Park et al., "Dual-Precision Deep Neural Network".
Park teaches train dual-precision DNNs that can switch between the two precision modes without the need for re-training. By sharing the common bits (b bits) between the low- and high-precision weights, the dual-precision DNN allows dynamic on-line switch between the two precision modes only with truncation (precision down-scaling) or concatenation (precision up-scaling) of the last 1 bits. However, Park does not teach a second machine learning model is trained by generating an adjustment loss that compares the outputs of three models for the same adjustment data exemplar: a baseline machine learning model, the first machine learning model, and the second machine learning model. The adjustment loss guides the second learning model to learn output corrections relative to the baseline while accounting for the behavior of the first machine learning model.
Kim et al., "Domain adaptation without source data."
Kim discusses source data free domain adaptation (SFDA), a domain adaption approach that eliminates the need to access source domain training data. Instead, it uses a pre trained source model and progressively updates a target model through self-learning. The method leverages low self-entropy target samples to correctly classified, to facilitate reliable adaption while preserving source domain privacy. However, Kim does not teach a second machine learning model is trained by generating an adjustment loss that compares the outputs of three models for the same adjustment data exemplar: a baseline machine learning model, the first machine learning model, and the second machine learning model. The adjustment loss guides the second learning model to learn output corrections relative to the baseline while accounting for the behavior of the first machine learning model.
Claim 26 recites:
wherein training the third machine learning model comprises generating an adjustment loss for a first block of the third plurality of blocks based on: (i) a first feature map generated by a first block of the first plurality of blocks based on a first exemplar in the adjustment data; (ii) a second feature map generated by a first block of the second plurality of blocks based on the first exemplar, wherein the first block from the second plurality of blocks comprises a quantized version of the first block from the first plurality of blocks; and (iii) a third feature map generated by the first block of the third plurality of blocks based on the first exemplar, wherein the first block of the third plurality of blocks corresponds to the first block of the second plurality of blocks.
Closest prior art:
Park et al., "Dual-Precision Deep Neural Network".
Park teaches train dual-precision DNNs that can switch between the two precision modes without the need for re-training. By sharing the common bits (b bits) between the low- and high-precision weights, the dual-precision DNN allows dynamic on-line switch between the two precision modes only with truncation (precision down-scaling) or concatenation (precision up-scaling) of the last 1 bits. However, Park does not teach a third machine learning model is trained by generating an adjustment loss that compares feature maps from corresponding first blocks of three models for the same adjustment data exemplar: the first machine learning model, a quantized model (second machine learning model), and third machine learning model. The adjustment loss enables the third machine learning model to learn feature level adjustments that compensate for errors introduced by quantization.
Kim et al., "Domain adaptation without source data."
Kim discusses source data free domain adaptation (SFDA), a domain adaption approach that eliminates the need to access source domain training data. Instead, it uses a pre trained source model and progressively updates a target model through self-learning. The method leverages low self-entropy target samples to correctly classified, to facilitate reliable adaption while preserving source domain privacy. However, Kim does not teach a third machine learning model is trained by generating an adjustment loss that compares feature maps from corresponding first blocks of three models for the same adjustment data exemplar: the first machine learning model, a quantized model (second machine learning model), and third machine learning model. The adjustment loss enables the third machine learning model to learn feature level adjustments that compensate for errors introduced by quantization.
Claim 27 recites:
wherein training the third machine learning model comprises generating an adjustment loss for a first block of the third plurality of blocks based on: (i) a first model output generated by the first machine learning model based on a first exemplar in the adjustment data; (ii) a second model output generated by the second machine learning model based on the first exemplar; and (iii) a third model output generated by the third machine learning model based on the first exemplar.
Closest prior art:
Park et al., "Dual-Precision Deep Neural Network".
Park teaches train dual-precision DNNs that can switch between the two precision modes without the need for re-training. By sharing the common bits (b bits) between the low- and high-precision weights, the dual-precision DNN allows dynamic on-line switch between the two precision modes only with truncation (precision down-scaling) or concatenation (precision up-scaling) of the last 1 bits. However, Park does not teach a third machine learning model is trained by generating an adjustment loss that compares the outputs of the first, second and third machine learning models for the same adjustment data exemplar. The adjustment loss is used to train the third machine learning model to adjust its outputs relative to those of the first and second machine learning model.
Kim et al., "Domain adaptation without source data."
Kim discusses source data free domain adaptation (SFDA), a domain adaption approach that eliminates the need to access source domain training data. Instead, it uses a pre trained source model and progressively updates a target model through self-learning. The method leverages low self-entropy target samples to correctly classified, to facilitate reliable adaption while preserving source domain privacy. However, Kim does not teach a third machine learning model is trained by generating an adjustment loss that compares the outputs of the first, second and third machine learning models for the same adjustment data exemplar. The adjustment loss is used to train the third machine learning model to adjust its outputs relative to those of the first and second machine learning model.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Sriram et al., Pub. No.: US20220044114A1.
Sriram teaches techniques to use low precision quantization to train a neural network, one or more weights of a trained model are represented by low bit integer numbers instead of using full floating point precision.
Sharma et al., Pub. No.: US20200226444A1.
Sharma describes a heterogenous architecture for training quantized neural networks is described, a hardware accelerator for training quantized data, comprises memory to store data, a plurality of compute units to perform computations of a data type for an inference phase of training quantized data of a neural network.
Any inquiry concerning this communication or earlier communications from the examiner
should be directed to MATIYAS T MARU whose telephone number is (571)270-0902 or via email: matiyas.maru@uspto.gov. The examiner can normally be reached Monday 8:00am - Friday 4:00pm 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 on (571)431-0762. The fax phone number for the organization were 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.
/M.T.M./ Examiner, Art Unit 2148
/MICHELLE T BECHTOLD/Supervisory Patent Examiner, Art Unit 2148
Prosecution Insights