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 .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 12/06/2023, 03/18/2025, and 05/14/2026 are in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statements are being considered by the examiner.
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 1-25 are rejected under 35 U.S.C. 101 for reciting an abstract idea without significantly more.
Regarding claim 1
Claim 1 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 1 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
“determining if an entry corresponding to the one or more inputs is stored in a memory lookup table for the neural network, wherein the memory lookup table comprises a plurality of entries each associating a respective neural network input with a corresponding output generated by a version of the neural network;” – this limitation recites determining whether received information corresponds to stored information in a table identifying an associated stored output. This is an evaluation or judgment that may be practically performed in the human mind, or with pen and paper, by comparing an input identifier or input representation to stored table entries and determining whether a corresponding output is associated with that entry. See MPEP § 2106.04(a)(2)(III).
Claim 1 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application. The additional elements:
“receiving one or more inputs for a neural network;” – the recited receiving step is mere data gathering under MPEP § 2106.05(g).
“responsive to the entry being stored in the memory lookup table, providing the corresponding output for the entry;” – this limitation is insignificant post-solution activity because it merely outputs/provides the result associated with the abstract determination once the stored entry is found. See MPEP § 2106.05(g).
“and responsive to the entry not being stored in the memory lookup table, providing an output by processing the one or more inputs by the neural network to generate the output.” – this limitation merely uses a neural network as a tool to generate an output when the stored table association is unavailable. The claim does not recite a specific neural-network architecture, particular training technique, particular layer arrangement, improved computer-memory structure, or other specific technical improvement. See MPEP §§ 2106.05(f), 2106.05(g), and 2106.05(h).
Claim 1 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception. The additional elements are:
“receiving one or more inputs for a neural network;” – this limitation is merely receiving one or more inputs for later use in the abstract determination. See MPEP § 2106.05(g).
“responsive to the entry being stored in the memory lookup table, providing the corresponding output for the entry;” – this limitation is merely receiving data for later us in the abstract determination. See MPEP § 2106.05(g).
“and responsive to the entry not being stored in the memory lookup table, providing an output by processing the one or more inputs by the neural network to generate the output.” – this limitation recites generic data processing using a neural network at a high level of generality. The claim does not recite unconventional hardware, a particular neural-network architecture, a specific improved computer implementation, or any non-generic arrangement of components. The neural network is used as a generic tool to process input data and generate output data. See MPEP §§ 2106.05(d) and 2106.05(f).
Regarding claim 2
Claim 2 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 2 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
“and comparing the signature for the one or more inputs to the memory lookup table.” – this limitation recites comparing generated information, i.e., a signature, to stored table information. Comparing one item of information to another item of store information is an evaluation or judgment that may be practically performed in the human mind, or with pen and paper, by comparing generated identifier/signature to stored table entries and determining whether a corresponding table entry exists. See MPEP § 2106.04(a)(2)(III).
Claim 2 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application. The additional elements:
“generating a signature for the one or more inputs by processing the one or more inputs by the neural network until a memory check layer of the neural network is reached;” – this limitation merely uses the neural network as a tool to generate intermediate information, i.e., a signature, for later use in the abstract comparison. The claim does not recite a specific neural-network architecture, a specific type of layer, a specific signature-generation algorithm, or a specific improvement to neural-network or computer functionality. Additionally, selecting a layer at which data is generated or checked does not, without more, improve the functioning of the computer or neural network, and does not effect a transformation, and does not impose a meaningful technological limit on the abstract idea. See MPEP §§ 2106.05(f), 2106.05(g), and 2106.05(h).
Claim 2 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception. The additional elements are:
“generating a signature for the one or more inputs by processing the one or more inputs by the neural network until a memory check layer of the neural network is reached;” – this limitation is recited at a high level of generality and merely uses a neural network to generate data for the abstract comparison. Generic computer implementation of processing input data to generate intermediate data is well-understood, routine, and conventional (WURC). See MPEP § 2106.05(d). The limitation also amounts to mere instructions to apply the abstract idea using a computer/neural network. See MPEP § 2106.05(f).
Regarding claim 3
Claim 3 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 3 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
“updating the signature for the one or more inputs as stored in the entry.” – this limitation recites updating stored information in a table entry. Updating stored information is a data-maintenance operation that may be practically performed in the human mind, or with pen and paper, by modifying a stored signature associated with a table entry. See MPEP § 2106.04(a)(2)(III).
Claim 3 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application.
Claim 3 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception.
Regarding claim 4
Claim 4 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 4 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
“flagging the one or more inputs for potential storage in the memory lookup table in response to an output threshold for the one or more inputs being satisfied.” – this limitation recites evaluating whether an output threshold is satisfied and, based on that evaluation, marking or identifying the inputs for potential storage. Evaluating whether a threshold is satisfied and flagging information based on that evaluation is an observation, evaluation, or judgment that may be practically performed in the human mind, or with pen and paper, by comparing an output value to a threshold and marking the corresponding input for later storage. See MPEP § 2106.04(a)(2)(III).
Claim 4 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application.
Claim 4 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception.
Regarding claim 5
Claim 5 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 5 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
“wherein comparing the signature for the one or more inputs to the memory lookup table is based on a tolerance range.” – this limitation recites applying a comparison rule when comparing information to stored table information. Comparing a signature to stored information using a tolerance range is an evaluation or judgment that may be practically performed in the human mind, or with pen and paper. See MPEP § 2106.04(a)(2)(III).
Claim 5 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application.
Claim 5 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception.
Regarding claim 6
Claim 6 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 6 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
“determining whether another one or more inputs is flagged for potential storage in the memory lookup table;” – this limitation recites determining whether information has been marked or identified for potential storage. Determining whether data has been flagged is an evaluation or judgment that may be practically performed in the human mind, or with pen and paper, by checking whether an item of information has been marked for possible storage. See MPEP § 2106.04(a)(2)(III).
“and creating a new entry in the memory lookup table for the other one or more inputs in response to an output accuracy prediction for at least one layer of the neural network up to the predetermined layer threshold meeting a layer output threshold.” – this limitation recites evaluating whether a prediction value satisfies a threshold and, based on that evaluation, adding information to a table. Comparing a prediction value to a threshold and recording an associated entry in a table is an observation, evaluation, or judgment that may be practically performed in the human mind, or with pen and paper. See MPEP § 2106.04(a)(2)(III).
Claim 6 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application. The additional elements:
“responsive to the other one or more inputs being flagged, processing the other one or more inputs by the neural network up to a predetermined layer threshold;” – this limitation merely uses a neural network as a tool to process input data up to a selected layer location. The claim does not recite a particular neural-network architecture, a specific layer structure, a specific technical basis for the layer threshold, or an improvement to neural-network or computer functionality. Rather, the limitation generally links the abstract data evaluation and storage process to a neural-network environment and specifies where processing stops for purposes of the abstract threshold evaluation. See MPEP §§ 2106.05(f) and 2106.05(h).
Claim 6 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception. The additional elements are:
“responsive to the other one or more inputs being flagged, processing the other one or more inputs by the neural network up to a predetermined layer threshold;” – this limitation is recited at a high level of generality and merely uses a neural network to process input data up to a layer. Generic computer implementation of processing data using a model, selecting at a layer, and generating intermediate information is well-understood, routine, and conventional (WURC). See MPEP § 2106.05(d). The limitation also amounts to no more than applying the abstract idea using a generic computer/neural-network tool. See MPEP § 2106.05(f).
Regarding claim 7
Claim 7 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 7 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
“wherein creating the new entry in the memory lookup table is further performed in response to a count associated with the one or more inputs meeting a count threshold.” – this limitation recites evaluating whether a count associated with information satisfies a threshold and, based on that evaluation, or judgment that may be practically performed in the human mind, or with pen and paper. See MPEP § 2106.04(a)(2)(III).
Claim 7 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application.
Claim 7 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception.
Regarding claim 8
Claim 8 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 8 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
Claim 8 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application. The additional elements:
“determining a memory check layer in the neural network.” – this limitation merely identifies where the abstract memory-lookup comparison is performed. The claim does not recite a particular neural-network architecture, a particular layer type, a specific technique for selecting the memory check layer, a particular hardware implementation, or any improvement to computer, memory, or neural-network functionality. Rather, the limitation functionally recites determining a layer for performing the abstract lookup-table check. See MPEP §§ 2106.04(d), 2106.05(a), 2106.05(f), and 2106.05(h).
Claim 8 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception. The additional elements are:
“determining a memory check layer in the neural network.” – this limitation is recited at a high level of generality and merely selects or identifies a neural-network layer at which the abstract comparison is performed. Generic computer implementation of selecting a processing point, identifying a layer, and performing data comparison at that point is well-understood, routine, and conventional (WURC). See MPEP § 2106.05(d). This limitation also amounts to no more than applying the abstract idea using a generic computer/neural-network tool. See MPEP § 2106.05(f).
Regarding claim 9
Claim 9 – Step 1 – Is the claim to a process, machine, manufacture or composition of matter?
Yes, the claim is to a process.
Claim 9 – Step 2A – Prong 1 – Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes, the claim recites an abstract idea.
Claim 9 – Step 2A – Prong 2 – Does the claim recite additional elements that integrate the judicial exception into a practical application?
No. There are no additional elements that integrate the judicial exception into a practical application. The additional elements:
“periodically retraining the neural network.” – this limitation merely recites, at a high level, updating the neural network over time. The claim does not specify how the retraining is technically performed, how the periodic interval is selected, how training data are selected or weighted, how the neural-network architecture is modified, or how the retraining improves computer functionality. Rather, the limitation generally links the abstract lookup-table process to generic neural-network maintenance. See MPEP §§ 2106.04(d), 2106.05(a), 2106.05(f), 2106.05(g), and 2106.05(h).
Claim 9 – Step 2B – Does the claim recite additional elements that amount to significantly more than the judicial exception?
No. There are no additional elements that amount to significantly more than the judicial exception. The additional elements are:
“periodically retraining the neural network.” – this limitation is recited at a high level of generality and merely updates a neural network at intervals using training. Generic computer implementation of retraining or updating a model using data is well-understood, routine, and conventional (WURC). See MPEP § 2106.05(d). The limitation also amounts to no more than applying the abstract lookup-table process using a generic neural-network tool. See MPEP § 2106.05(f).
Regarding claims 10-18
Claims 10-18 are apparatus counterparts to method claims 1-9. Claim 10 recites a generic processing device and memory storing instructions that cause the processing device to perform substantially the same operations recited in claim 1, and claims 11-18 correspond to the additional limitations of claims 2-9.
Under Step 2A, Prong One, claims 10-18 recite the same abstract idea identified for claims 1-9. In particular, claim 10 recites determining whether an entry corresponding to receiving inputs is stored in a memory lookup table, where entries associate neural-network inputs with corresponding outputs. This is a mental process because it involves evaluating whether information corresponds to stored table information and identifying an associated output. See MPEP § 2106.04(a)(2)(III). Claims 11-18 add the same abstract evaluations discussed for claims 2-9, including comparing signatures, updating stored information, applying thresholds, creating entries, selecting a check layer, and periodically retraining a model.
Under Step 2A, Prong Two, the additional apparatus elements do not integrate the abstract idea into a practical application. The recited processing device and memory merely implement the abstract idea using generic computer components. The claims do not recite a particular processor, particular memory architecture, specific improved neural-network architecture, or other improvement to computer functionality. The receiving/providing steps remain insignificant data-gathering or output activity under MPEP § 2106.05(g), and the generic computer implementation merely applies the abstract idea using a computer under MPEP § 2106.05(f). The claims also generally link the abstract idea to a neural-network environment without imposing a meaningful technological limit. See MPEP § 2106.05(h).
Under Step 2B, the additional elements, individually and as an ordered combination, do not amount to significantly more. Generic processing devices and memory storing instructions for receiving, storing, comparing, retrieving, and outputting data are well-understood, routine, and conventional (WURC) computer components/functions. See MPEP § 2106.05(d). The claims merely implement the ineligible method operations of claims 1-9 on generic apparatus components and therefore do not add an inventive concept.
Accordingly, claims 10-18 are ineligible under 35 U.S.C. § 101.
Regarding claims 19-25
Claims 19-25 are computer program-product counterparts to method claims 1-7. Claim 19 recites a generic computer readable storage medium storing instructions that cause the processing device to perform substantially the same operations recited in claim 1, and claims 20-25 correspond to the additional limitations of claims 2-7.
Under Step 2A, Prong One, claims 19-25 recite the same abstract idea identified for claims 1-7. In particular, claim 19 recites determining whether an entry corresponding to receiving inputs is stored in a memory lookup table, where entries associate neural-network inputs with corresponding outputs. This is a mental process because it involves evaluating whether information corresponds to stored table information and identifying an associated output. See MPEP § 2106.04(a)(2)(III). Claims 20-25 add the same abstract evaluations discussed for claims 2-7, including comparing signatures, updating stored information, applying thresholds, creating entries, selecting a check layer, and periodically retraining a model.
Under Step 2A, Prong Two, the additional apparatus elements do not integrate the abstract idea into a practical application. The recited processing device and memory merely implement the abstract idea using generic computer components. The claims do not recite a particular processor, particular memory architecture, specific improved neural-network architecture, or other improvement to computer functionality. The receiving/providing steps remain insignificant data-gathering or output activity under MPEP § 2106.05(g), and the generic computer implementation merely applies the abstract idea using a computer under MPEP § 2106.05(f). The claims also generally link the abstract idea to a neural-network environment without imposing a meaningful technological limit. See MPEP § 2106.05(h).
Under Step 2B, the additional elements, individually and as an ordered combination, do not amount to significantly more. Generic processing devices and memory storing instructions for receiving, storing, comparing, retrieving, and outputting data are well-understood, routine, and conventional (WURC) computer components/functions. See MPEP § 2106.05(d). The claims merely implement the ineligible method operations of claims 1-7 on generic apparatus components and therefore do not add an inventive concept.
Accordingly, claims 19-25 are ineligible under 35 U.S.C. § 101.
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) 1-6, 8, 10-15, 17, and 19-24 is/are rejected under 35 U.S.C. 103 as being unpatentable over Xiaohu Tang (LUT-NN: Empower Efficient Neural Network Inference with Centroid Learning and Table Lookup) in view of Alperen Gormez (
E
2
CM: Early Exit via Class Means for Efficient Supervised and Unsupervised Learning).
Regarding claim 1, Tang in view of Gormez teach a method comprising:
“receiving one or more inputs for a neural network;” – Tang teaches this limitation. Tang teaches DNN inference using input features of DNN layers:
“Each layer of a DNN model is to output another level of features given the input features.” (Tang, § 1 Introduction)
Tang further states that, during inference:
“the results of the closest centroids with the input features can be read directly from the table as the approximated output of this operator.” (Tang, § 1 Introduction)
Tang also describes the inference architecture as operating on input tensors:
“LUT-NN model inference architecture, comprising the Closest Centroid Search stage and the Table Read and Accumulation stage. In the Closest Centroid Search stage, LUT-NN first computes the distance between input tensors and centroids and searches the nearest centroids for each input tensor.” (Tang, § 5 Table Lookup Inference Design)
“determining if an entry corresponding to the one or more inputs is stored in a memory lookup table for the neural network,” – Tang teaches determining the closest stored centroid entry corresponding to input features of a neural network:
“The system learns the typical features, named centroid, for each linear computation operator, and precompute the results for these centroids to save in lookup tables.” (Tang, § 1 Introduction)
Tang further states:
“During inference, the results of the closest centroids with the input features can be read directly from the table as the approximated output of this operator.” (Tang, § 1 Introduction)
Tang expressly describes the determination step as a closest-centroid search:
“In the Closest Centroid Search stage, LUT-NN first computes the distance between input tensors and centroids and searches the nearest centroids for each input tensor.” (Tang, § 5 Table Lookup Inference Design)
Tang further states:
“after computing centroid distances, the Closest Centroid Search stage must identify the nearest centroid for each input sub-vector and generate the centroid index” (Tang, § 5.1 Closest centroid search)
Tang teaches that the centroid index is used to access the memory lookup table:
“LUT-NN obtains the indices of the nearest centroids for the input sub-vectors after closest centroid search, and leverages Table Read and Accumulation stage to compute the final results.” (Tang, § 5.2 Table read and accumulation)
“wherein the memory lookup table comprises a plurality of entries each associating a respective neural network input with a corresponding output generated by a version of the neural network;” – Tang teaches this limitation. Tang teaches that the lookup table stores precomputed results for centroids, where the centroids are representative neural-network input features:
“The system learns the typical features, named centroid, for each linear computation operator, and precompute the results for these centroids to save in lookup tables.” (Tang, § 1 Introduction)
Tang further explains the input/output association:
“the typical features can be learned for each computation operator of DNNs, so that the output of these typical features can represent the output of diverse features.” (Tang, § 1 Introduction)
Tang then states:
“by precomputing and saving the output of typical features, the output of future inputs can be read directly without computation.” (Tang, § 1. Introduction)
Tang also teaches construction of the lookup table by precomputing neural-network operator results:
“Since B is constant, the multiplication of all the centroids and B can be precomputed to construct a lookup table” (Tang, § 2.2 PQ for AMM)
Tang further states:
“The matrix multiplication can then be approximated by looking up and aggregating the results of the nearest centroids in the precompute table” (Tang, § 2.2 PQ for AMM)
Thus, Tang teaches plural lookup-table entries in which each centroid / typical neural-network input feature is associated with a corresponding precomputed operator output generated from the DNN operator weights, i.e., by a version of the neural network.
“responsive to the entry being stored in the memory lookup table, providing the corresponding output for the entry;” – Tang teaches this limitation. Tang teaches that once the closest centroid entry is identified, the corresponding precomputed result is read from the lookup table as the output:
“During inference, the results of the closest centroids with the inputs can be read directly from the table, as the approximated outputs without computations.” (Tang, § Abstract)
Tang similarly states
“During inference, the results of the closest centroids with the input features can be read directly from the table as the approximated output of this operator.” (Tang, § 1 Introduction)
Tang further teaches:
“This stage first reads the pre-computed results from the corresponding lookup table through indices … and completes the computation by accumulation operation” (Tang, § 5.2 Table read and accumulation)
“and responsive to the entry not being stored in the memory lookup table, providing an output by processing the one or more inputs by the neural network to generate the output.” – Tang teaches this limitation. Tang teaches that LUT-NN does not necessarily replace all neural-network processing with table lookup:
“For operators to replace, we replace all convolution operators for CNN models by table lookup except the first one.” (Tang, § 6.1 Experiment methodologies and settings)
Tang further states:
“For BERT, other than explicitly stated, we replace the fully connected operators of the last 6 layers.” (Tang, § 6.1 Experiment methodologies and settings)
Thus, Tang teaches a neural-network inference arrangement in which some operators are serviced by lookup table and other operators/layers remain processing by the neural network to generate the output.
To the extent Tang is considered not to expressly teach the claimed response to the lookup-table entry not being stored, Gormez teaches continuing neural-network processing when the stored representative / early-exit condition is not satisfied:
“During inference, the decision of exiting after
l
j
or moving forward to
l
j
+
1
is made according to a threshold value
T
j
.” (Gormez, § III. Early Exit Class Means)
Gormez further states:
“Otherwise, the input moves forward to the next layer. In the worst case, execution ends at the last layer of the network.” (Gormez, § III. Early Exit Class Means)
Accordingly, Tang teaches selective use of lookup-table inference within a neural network, and Gormez teaches the fallback behavior of continuing neural-network processing when a stored representative match / early-exit condition is not satisfied.
It would have been obvious to a person of ordinary skill in the art to incorporate Gormez’s fallback / continued-processing behavior into Tang’s lookup-table DNN inference framework so that, when a lookup-table condition is not satisfied, the neural network continues processing the input to generate an output.
Tang expressly motivates replacing DNN computation with table lookup to reduce inference cost. Gormez likewise motivates avoiding unnecessary neural-network computation while continuing neural-network processing when early exit is not appropriate.
A POSITA would have been motivated to use Gormez’s known continue-processing fallback in Tang’s table-lookup inference system to ensure an output is generated when lookup-table substitution is not applicable, with the predictable result of retaining neural-network accuracy while reducing inference cost where table-lookup entries are applicable.
Regarding claim 2, Tang in view of Gormez teach the method of claim 1, wherein determining if an entry corresponding to the one or more inputs is stored in the memory lookup table comprises:
“and comparing the signature for the one or more inputs to the memory lookup table.” – Tang teaches this limitation. Tang teaches comparing input tensors/sub-vectors to stored centroids associated with lookup-table entries:
“In the Closest Centroid Search stage, LUT-NN first computes the distance between input tensors and centroids and searches the nearest centroids for each input tensor.” (Tang, § 5 Table Lookup Inference Design)
Tang further states:
“after computing centroid distances, the Closest Centroid Search stage must identify the nearest centroid for each input sub-vector and generate the centroid index” (Tang, § 5.1 Closest centroid search)
Tang teaches that the nearest-centroid index is used for lookup-table access:
“This stage first reads the pre-computed results from the corresponding lookup table through indices … and completes the computation by accumulation operation” (Tang, § 5.2 Table read and accumulation)
Tang does not teach this limitation:
“generating a signature for the one or more inputs by processing the one or more inputs by the neural network until a memory check layer of the neural network is reached;”
Gormez, however, teaches this limitation:
“generating a signature for the one or more inputs by processing the one or more inputs by the neural network until a memory check layer of the neural network is reached;” – Gormez teaches processing the input through the neural-network layers and using a layer output as the representation/signature for comparison at a potential exit layer. Gormez defines the network as:
“We denote the network
F
with
M
layers as a sequence
l
1
,
l
2
,
…
,
l
M
.” (Gormez, § III. Early Exit Class Means)
And states the layer processing relationship:
“
x
j
(
i
)
=
l
j
x
j
-
1
i
,
j
=
1,2
,
…
,
M
.
” (Gormez, § III. Early Exit Class Means)
Gormez teaches that the layer output is used for comparison against stored class means:
“During inference, output of a layer is compared with the corresponding class means using Euclidean distance as the metric.” (Gormez, § 1. Introduction)
Gormez further identifies the layer at which the check is made:
“During inference, the decision of exiting after
l
j
or moving forward to
l
j
+
1
is made according to a threshold value
T
j
.” (Gormez, § 1. Introduction)
Thus, Gormez teaches processing the input through the neural network until a check layer
l
j
is reached, where the layer output
x
j
(
i
)
functions as the claimed signature.
It would have been obvious to a person of ordinary skill in the art to use Gormez’s layer-output early-exit check as the “memory check layer” / signature-generation step in Tang’s lookup-table inference framework. Tang already teaches finding corresponding lookup-table entries by computing distance between input tensors and stored centroids, while Gormez teaches performing a similar distance-based check at an intermediate neural-network layer and deciding whether to exit or continue. The combination would predictably allow the lookup-table determination of claim 1 to be performed using a signature generated at a neural-network layer, reducing inference cost while preserving continued neural-network processing when the check is not satisfied.
Regarding claim 3, Tang in view of Gormez teach the method of claim 2, further comprising
“updating the signature for the one or more inputs as stored in the entry.” – Tang teaches this limitation. Tang teaches that LUT-NN learns centroid entries representing typical input features of neural-network operators:
“The system learns the typical features, named centroid, for each linear computation operator, and precompute the results for these centroids to save in lookup tables.” (Tang, § 1 Introduction)
Tang teaches updating those stored centroid/signature values through backpropagation and gradient descent:
“the key factor for DNN centroid learning is to pass the model loss to each operator through backpropagation, and iteratively adjust the centroids by the gradients to minimize the model loss.” (Tang, § 1 Introduction)
Tang further states:
“Through back propagation, it adapts three different levels of approximation introduced by centroids to model inference by three methods, and minimize the accuracy loss.” (Tang, § 1 Introduction)
Regarding claim 4, Tang in view of Gormez teach the method of claim 1, further comprising
“” – Tang teaches this limitation in part. Tang teaches that input features are learned as centroids and used to build lookup tables:
“LUT-NN learns the typical features for each operator, named centroid, and precompute the results for these centroids to save in lookup tables.” (Tang, § Abstract)
Tang further teaches that the lookup-table entries correspond to learned centroids and that future inputs use those entries during inference:
“During inference, the results of the closest centroids with the inputs can be read directly from the table, as the approximated outputs without computations.” (Tang, § Abstract)
Tang also teaches collecting operator inputs from training data and using them to initialize centroid entries:
“we apply the original model inference to a randomly sampled sub-dataset (consisting of 1024 training samples) and collect each operator’s inputs. We then utilize the k-means algorithm to cluster each operator’s inputs and obtain the initial centroids.” (Tang, § 6.1 Experiment methodologies and settings)
Thus, Tang teaches selecting/identifying neural-network input features for potential storage in look-up table form as centroids and corresponding precomputed results.
Tang does not teach these limitations:
“flagging … in response to an output threshold for the one or more inputs being satisfied.”
Gormez, however, teaches these limitations:
“flagging … in response to an output threshold for the one or more inputs being satisfied.” – Gormez teaches:
“During inference, the decision of exiting after
l
j
or moving forward to
l
j
+
1
is made according to a threshold value
T
j
.” (Gormez, § III. Early Exit Class Means)
Gormez further states:
“If the largest softmax probability is greater than the specified threshold
T
J
, execution is stopped and the class with the largest softmax probability is predicted.” (Gormez, § III. Early Exit Class Means)
The largest softmax probability is an output confidence/probability for the input, and the specified
T
J
corresponds to the claimed output threshold.
It would have been obvious to a person of ordinary skill in the art to flag or select inputs satisfying Gormez’s high-confidence output threshold for potential storage as Tang-style lookup-table entries. Tang teaches building lookup-table entries from typical neural-network input features to reduce inference cost, while Gormez teaches identifying inputs/layer outputs that satisfy a confidence threshold suitable for early prediction. Combining these teachings would predictably store or consider high-confidence representative inputs for lookup-table reuse, reducing later neural-network computation while preserving inference reliability.
Regarding claim 5, Tang in view of Gormez, teach the method of claim 2, wherein
“comparing the signature for the one or more inputs to the memory lookup table ” – Tang teaches this limitation in part. Tang teaches comparing input tensors/sub-vectors to stored centroids and identifying the nearest centroid index:
“In the Closest Centroid Search stage, LUT-NN first computes the distance between input tensors and centroids and searches the nearest centroids for each input tensor.” (Tang, § 5 Table Lookup Inference Design)
Tang further teaches that, after distance computation:
“after computing centroid distances, the Closest Centroid Search stage must identify the nearest centroid for each input sub-vector and generate the centroid index” (Tang, § 5.1 Closest centroid search)
Tang also teaches that the centroid index is used to access the lookup table:
“LUT-NN obtains the indices of the nearest centroids for the input sub-vectors after closest centroid search, and leverages Table Read and Accumulation stage to compute the final results.” (Tang, § 5.2 Table read and accumulation)
Tang does not teach this portion of the limitation:
“… is based on a tolerance range.”
Gormez, however, teaches this remaining portion of the limitation:
“… is based on a tolerance range.” – Gormez teaches applying a threshold value to determine whether the layer-output signature is sufficiently close to stored class-mean information. Gormez discloses comparing a neural-network layer output to stored representative values using distance:
“The Euclidean distance between a layer output
x
j
(
i
)
and K class means
c
j
k
is then computed via
d
j
k
(
i
)
=
x
j
(
i
)
-
c
j
k
2
” (Gormez, § III. Early Exit Class Means)
Gormez further teaches that the decision is based on a threshold:
“During inference, the decision of exiting after
l
j
or moving forward to
l
j
+
1
is made according to a threshold value
T
j
.” (Gormez, § III. Early Exit Class Means)
Gormez also states:
“If the largest softmax probability is greater than the specified threshold
T
J
, execution is stopped and the class with the largest softmax probability is predicted.” (Gormez, § III. Early Exit Class Means)
The claimed “tolerance range” is met by Gormez’s threshold-based distance/probability comparison, because the signature is accepted for early output only when it falls within the threshold condition corresponding to sufficient closeness/confidence.
Tang teaches comparing an input-derived signature to lookup-table entries by computing distances between input tensors/sub-vectors and stored centroids and generating a nearest-centroid index for table lookup. Gormez teaches that such comparison may be based on a tolerance range because it computes Euclidean distance between a layer-output signature and stored class means and determines whether to exit according to a threshold value
T
J
. It would have been obvious to apply Gormez’s threshold/tolerance-based comparison to Tang’s nearest-centroid lookup-table comparison to control when a stored entry is sufficiently for lookup-table reuse, thereby reducing computation while maintaining inference accuracy.
Regarding claim 6, Tang in view of Gormez, teach the method of claim 1, further comprising:
“” – Tang teaches this limitation in part. Tang teaches that LUT-NN stores representative neural-network input features as centroids and corresponding precomputed outputs in lookup tables:
“LUT-NN learns the typical features for each operator, named centroid, and precompute the results for these centroids to save in lookup tables.” (Tang, § Abstract)
Tang further teaches collecting inputs for potential centroid/lookup-table construction:
“we apply the original model inference to a randomly sampled sub-dataset (consisting of 1024 training samples) and collect each operator’s inputs. We then utilize the k-means algorithm to cluster each operator’s inputs and obtain the initial centroids.” (Tang, § 6.1 Experiment methodologies and settings)
Thus, Tang teaches inputs considered for storage in lookup tables.
“responsive to the other one or more inputs being flagged, ” – Tang teaches this limitation in part. Tang teaches selective processing / replacement of particular layers or operators:
“For operators to replace, we replace all convolution operators for CNN models by table lookup except the first one.” (Tang, § 6.1 Experiment methodologies and settings)
This confirms that known lookup-table inference systems selectively process or replace neural-network operations up to or at selected layers/operators.
“and creating a new entry in the memory lookup table for the other one or more inputs ” – Tang teaches this limitation in part. Tang teaches creating lookup-table entries by learning centroids from neural-network inputs and precomputing/storing their corresponding results:
“The system learns the typical features, named centroid, for each linear computation operator, and precompute the results for these centroids to save in lookup tables.” (Tang, § 1 Introduction)
Tang further teaches table construction from centroids and weights:
“Since
B
is constant, the multiplication of all the centroids and
B
can be precomputed to construct a lookup table” (Tang, § 1 Introduction)
Tang also states:
“The matrix multiplication can then be approximated by looking up and aggregating the results of the nearest centroids in the precompute table” (Tang, § 1 Introduction)
Thus, Tang teaches creating new lookup-table entries for representative input features by storing precomputed centroid/operator results.
Tang does not teach these limitations and/or portions of:
“determining whether … is flagged …”
“… processing the other one or more inputs by the neural network up to a predetermined layer threshold;”
“… in response to an output accuracy prediction for at least one layer of the neural network up to the predetermined layer threshold meeting a layer output threshold.”
Gormez, however, teaches these limitations and/or portions of:
“determining whether … is flagged …” – Gormez teaches evaluating a received input at a layer and determining whether a threshold is satisfied:
“During inference, the decision of exiting after or moving forward to
l
j
+
1
is made according to a threshold value
T
j
. If the largest softmax probability is greater than the specified threshold
T
j
, execution is stopped and the class with the largest softmax probability is predicted.” (Gormez, § III. Early Exit Class Means)
Thus, Gormez teaches determining whether an input satisfies a threshold condition indicative of being selected/flagged.
“… processing the other one or more inputs by the neural network up to a predetermined layer threshold;” – Gormez teaches processing an input through a neural network layer-by-layer and making decisions at specified layers:
“We denote the network
F
with
M
layers as a sequence
l
1
,
l
2
,
…
,
l
M
.” (Gormez, § III. Early Exit Class Means)
Gormez further states:
“The input-output relationships of the network F can be expressed as
x
j
(
i
)
=
l
j
(
x
j
-
1
(
i
)
)
,
j
=
1,2
,
…
,
M
.
” (Gormez, § III. Early Exit Class Means)
Gormez also teaches using thresholds associated with layers and target computation budget:
“The
j
t
h
computer of the threshold vector corresponds to
T
j
in Algorithm 1 and is utilized at layer
l
j
of the neural network.” (Gormez, § IV. Results)
Gormez therefore teaches processing an input through the neural network to a predetermined layer
l
j
having a threshold
T
j
, i.e., the claimed predetermined layer threshold.
“… in response to an output accuracy prediction for at least one layer of the neural network up to the predetermined layer threshold meeting a layer output threshold.” – Gormez teaches generating a prediction at a layer and comparing that prediction to a threshold:
“Let
y
^
(
i
)
denote the prediction of the network,
x
j
(
i
)
denote the output of layer
j
, and
y
^
j
(
i
)
denote the prediction in case the input exits the network after layer
j
,” (Gormez, § III. Early Exit Class Means)
Gormez further teaches converting distances at the layer to prediction probabilities:
“Finally, the normalized distances are converted to probabilities of input belonging to a class in order to perform inference. This is done using the softmax function as
P
y
^
j
i
=
k
=
s
o
f
t
m
a
x
(
-
d
j
i
)
.” (Gormez, § III. Early Exit Class Means)
Gormez then applies a layer threshold:
“During inference, the decision of exiting after or moving forward to
l
j
+
1
is made according to a threshold value
T
j
. If the largest softmax probability is greater than the specified threshold
T
j
, execution is stopped and the class with the largest softmax probability is predicted.” (Gormez, § III. Early Exit Class Means)
Thus, Gormez teaches a layer output prediction/probability for at least one neural-network layer meeting a layer-specific threshold
T
j
.
Tang teaches creating memory-lookup-table entries from neural-network input features by collecting operator inputs, clustering them into centroids, precomputing results for those centroids, ad saving the precomputed results in lookup tables. Gormez teaches determining, at a neural-network layer, whether an input’s layer-level prediction/probability satisfies a threshold
T
j
, and processing the input through the neural network layer-by-layer to such a threshold layer. It would have been obvious to a person of ordinary skill in the art to create Tang-style lookup-table entries for inputs or representative input features that satisfy Gormez’s layer-level prediction threshold, because both references seek to reduce neural-network inference cost while preserving acceptable prediction accuracy.
Regarding claim 8, Tang in view of Gormez, teach the method of claim 6, wherein the method of claim 6, further comprising
“determining a memory check layer in the neural network.” – Tang does not teach this limitation. Gormez, however, teaches this limitation. Gormez teaches performing an early-exit check at neural-network layer
l
j
, where a layer threshold
T
j
is used to determine whether to exit or continue:
“During inference, the decision of exiting after
l
j
or moving forward to
l
j
+
1
is made according to a threshold value
T
j
.” (Gormez, § III. Early Exit Class Means)
Gormez further teaches that thresholds are selected for layers and correspond to layer positions in the neural network:
“The
j
t
h
component of the threshold vector corresponds to
T
j
in Algorithm 1 and is utilized at layer
l
j
of the neural network.” (Gormez, IV. Results)
Gormez also teaches using thresholds to adjust the model according to computational needs:
“Also, thresholds make it possible to adjust the model according to various computational needs.” (Gormez, § IV. Results)
Thus, Gormez teaches determining a layer
l
j
of the neural network at which a threshold-based check is made, i.e., a memory check layer in the neural network.
It would have been obvious to determine a memory check layer in Tang’s lookup-table neural-network inference system using Gormez’s layer-specific threshold teachings because Tang already teaches selectively replacing only certain neural-network operators/layers with table lookup, and Gormez teaches making an exit/continue decision at selected layers using threshold values. A POSITA would have been motivated to identify the layer at which lookup-table comparison is performed in order to balance inference-cost reduction against accuracy, as both Tang and Gormez expressly address reducing neural-network inference cost while maintaining acceptable prediction accuracy.
Claims 7, 16, and 25 are rejected under 35 U.S.C. 103 as being unpatentable over Tang in view of Gormez and further in view of Yun Li (Boosting Mobile CNN Inference through Semantic Memory).
Regarding claim 7, Tang in view of Gormez and further in view of Li, teach the method of claim 6, wherein
“creating the new entry in the memory lookup table is further performed ” – Tang teaches this limitation in part. Tang teaches creating lookup-table entries by learning centroids from neural-network inputs and precomputing/storing their corresponding results:
“The system learns the typical features, named centroid, for each linear computation operator, and precompute the results for these centroids to save in lookup tables.” (Tang, § 1 Introduction)
Tang further teaches table construction from centroid and weights:
“Since B is constant, the multiplication of all the centroids and B can be precomputed to construct a lookup table” (Tang, § 2.2 PQ for AMM)
Tang also states:
“The matrix multiplication can then be approximated by looking up and aggregating the results of the nearest centroids in the precompute table” (Tang, § 2.2 PQ for AMM)
Thus, Tang teaches creating lookup-table entries for representative neural-network input features by storing recomputed centroid/operator results.
Tang does not teach this portion of the limitation:
“… in response to a count associated with the one or more inputs meeting a count threshold.”
Li, however, teaches this remaining portion of the limitation:
“… in response to a count associated with the one or more inputs meeting a count threshold.” – Li teaches maintaining a frequency/count table for observed input objects/classes and using that count-based information to select objects for caching in fast memory:
“SMTM employs a hierarchical memory architecture to leverage the long-tail distribution of objects of interest” (Li, § Abstract)
Li further teaches that SMTM uses memory based on frequently observed inputs:
“MTM promotes the most frequently- and recently-seen objects in the fast memory” (Li, § 1 Introduction)
Li further states:
“A frequency table and a time-tamp table are maintained in the memory replacement policy, which is used to update the fast memory periodically.” (Li, § 6 Adaptive Priming Memory)
Li further states:
“Frequency table keeps a record of the number of times that each object class presented in history. It is initialized as zeros and updated by every inference output during runtime.” (Li, § 6.1 Cache replacement policy)
Li also teaches selecting objects to be cached based on a count/frequency-based score:
“The replacement policy takes the Top-k highest score that are calculated by the following equation to select the objects from the global memory and cache them in the fast memory,” (Li, § 6.1 Cache replacement policy)
Lu further explains that the score includes frequency:
“where
F
T
i
is the frequency of object
i
in the frequency table …” (Li, § 6.1 Cache replacement policy)
Thus, Li teaches using a count associated with inputs/object classes, namely the number of times each object class has appeared, and selecting/cache-promoting entries based on a count/frequency-based score meeting a selection criterion, i.e., a count threshold.
It would have been obvious to a person of ordinary skill in the art to apply Li’s count/frequency-based cache admission or promotion policy to Tang’s neural-network lookup-table memory so that new entries are created or promoted for inputs that are repeatedly observed. Tang teaches creating lookup-table entries for neural-network inferences to reduce computation, and Li teaches that CNN inference can be accelerated by caching frequently and recently seen objects in fast memory using a frequency table. A POSITA would have been motivated to use Li’s frequency/count-based selection with Tang’s lookup-table entries to conserve limited memory, avoid storing low-reuse entries, and improve lookup hit rate for inputs likely to recur.
Claims 9 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Tang in view of Gormez and further in view of Rejith George Joseph (US10650432B1).
Regarding claim 9, Tang in view of Gormez and further in view of Joseph, teach the method of claim 1, further comprising
“periodically retraining the neural network.” – Tang does not teach this limitation. Joseph, however, teaches this limitation. Joseph teaches that a neural network may be retrained at predetermined intervals using updated data:
“The neural network can be trained at predetermined intervals using updated item interaction data. For example, the neural network can be trained once per day using purchase histories of users of an electronic commerce system.” (Joesph, col. 4, lines 8-12)
Joseph further teaches periodic retraining:
“The periodic retraining of the neural network, using updated data and the sliding windows created by the date split, allows the model to adapt to seasonal trends when predicting item interaction events.” (Joseph, col. 4, lines 30-33)
Joesph also states:
“According to some aspects, a neural network of a particular topology can be trained intermittently and/or periodically based on updated historical item interaction data.”
Thus, Joseph teaches periodically retraining the neural network.
It would have been obvious to a POSITA to periodically retrain the neural network of Tang in view of Gormez and Joseph in order to update the model using updated data and maintain prediction accuracy over time. Tang teaches a lookup-table neural-network inference arrangement whose accuracy is improved through training and iterative updating, Gormez teaches fine tuning / retraining neural-network models using local datasets, and Joseph teaches periodically retraining a neural network using updated data so the model adapts to changing trends. Therefore, Tang in view of Gormez and further in view of Joesph teaches periodically retraining the neural network.
Regarding claims 10-18 and 19-25, the apparatus and computer program product claims are rejected on the same grounds as their corresponding method claims.
Claims 10-18 recite apparatus limitations in which a processing device and memory store computer program instructions that cause the processing device to perform substantially the same operations recited in method claims 1-9. Claims 19-25 recite a computer program product comprising a computer readable storage medium storing computer program instructions that, when executed, perform substantially the same operations recited in method claims 1-7.
Tang teaches that LUT-NN is a computer-implemented neural network inference system that is implemented on conventional processing hardware:
“LUT-NN is implemented on ARM Neon and Intel SIMD ISA (Instruction Set Architecture), and can run in a single or multi threads” (Tang, § 6.1 Experiment methodologies and settings)
Tang further teaches a neural-network inference architecture using lookup-table execution:
“LUT-NN model inference architecture, comprising the Closest Centroid Search stage and the Table Read and Accumulation stage.” (Tang, § 5 Table Lookup Inference Design)
Joseph further confirms that machine-learning/neural-network processes are conventionally implemented using processors, memory, storage, and computer-readable media:
“The interactive computing system 500 may include at least one memory 506 and one or more processing units (or processor(s)) 508.” (Joseph, col. 13, lines 65-67)
“The memory 506 and the additional storage 512, both removable and non-removable, are examples of computer-readable storage media.” (Joseph, col. 14, lines 58-60)
Expressing the same computer-implemented operations as stored instructions in memory, or as instructions on a computer-readable storage medium, does not, without more, impart a patentable distinction over the previously established method combination. See MPEP §§ 2113 and 2114.
Accordingly:
Claims 10 and 19 are obvious over Tang in view of Gormez for the same reasons set forth with respect to claim 1.
Claims 11 and 20 are obvious over Tang in view of Gormez for the same reasons set forth with respect to claim 2.
Claims 12 and 21 are obvious over Tang in view of Gormez for the same reasons set forth with respect to claim 3.
Claims 13 and 22 are obvious over Tang in view of Gormez for the same reasons set forth with respect to claim 4.
Claims 14 and 23 are obvious over Tang in view of Gormez for the same reasons set forth with respect to claim 5.
Claims 15 and 24 are obvious over Tang in view of Gormez for the same reasons set forth with respect to claim 6.
Claims 16 and 25 are obvious over Tang in view of Gormez, and further in view of Li, for the same reasons set forth with respect to claim 7. Li teaches maintaining a frequency/count table and selecting objects for fast-memory caching based on frequency/recency-based scores.
Claim 17 is obvious over Tang in view of Gormez for the same reasons set forth with respect to claim 8.
Claims 18 is obvious over Tang in view of Gormez, and further in view of Joseph, for the same reasons set forth with respect to claim 9. Joesph teaches that “the neural network can be trained at predetermined intervals using updated item interaction data” and that “periodic retraining of the neural network” allows the model to adapt to updated trends (Joseph, col. 4, lines 8-9, 30-31)
Thus, claims 10-18 and 19-25 are unpatentable for the same reasons discussed for their corresponding method claims.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to PAUL COLEMAN whose telephone number is (571)272-4687. The examiner can normally be reached Mon-Fri.
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, David Yi can be reached at (571) 270-7519. 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.
/PAUL COLEMAN/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights