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 .
General Remarks
This communication is considered fully responsive to Applicant’s application filed 02/16/2026.
Application filed: 02/07/2023
Application PgPUB: 2023/0275812
Claims:
Claims 1, 3, 4, 6-8, 10, 11, 15-18 and 21-28 are pending.
Claims 1, 15, 21 and 26 are independent.
Claims 2, 5, 9, 12-14, 19 and 20 are canceled.
Claims 1, 15, 21, 26 and 27 are amended.
IDS:
Previous IDS:
IDS filed 02/07/2023 has been considered.
Continuity/Priority Data:
This Application claims priority to European Patent Application No. EP20305921.7 filed 08/10/2020.
This Application claims priority to European Patent Application No. EP20305922.5 filed 08/10/2020.
This Application is the 371 National Stage entry of International Patent Application No. PCT/EP2021/069944.
Response to Arguments
Applicant’s arguments, see Applicant's response, filed 02/16/2026, with respect to the rejection(s) of claim(s) 1, 3, 4, 6-8, 10, 11, 15-18 and 21-28 under 35 U.S.C. 102 and 103 have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of U.S. Patent Application Publication No. 2021/0160296 A1 to Lee et al. (“Lee”).
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 may not be obtained though the invention is not identically disclosed or described as set forth in section 102 of this title, if the differences between the subject matter sought to be patented and the prior art are such that the subject matter as a whole would have been obvious at the time the invention was made to a person having ordinary skill in the art to which said subject matter pertains. Patentability shall not be negatived by the manner in which the invention was made.
Claims 1, 3, 4, 6-8, 10 and 21-24 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication No. 2022/0283820 A1 to Pudipeddi et al. (“Pudipeddi”) in view of Printed Publication, “Distributing Deep Neural Networks with Containerized Partitions at the Edge” to Zhou et al. (“Zhou”) in further view of U.S. Patent Application Publication No. 2021/0232399 A1 to Yang et al. (“Yang”) in further view of U.S. Patent Application Publication No. 2021/0160296 A1 to Lee et al. (“Lee”).
As to claim 1, Pudipeddi discloses:
a method implemented by a server, comprising:
receiving information indicating a request for an artificial intelligence and/or machine learning (AI/ML) model, wherein said information is received from a client device via any of wired and wireless communication (Fig. 17, network, ¶0049 – Pudipeddi teaches At step 202, a portion of an artificial intelligence (AI) model is downloaded into memory of a target device from a parameter server that stores a master copy of the AI model. For example, target device 134a of FIG. 1, specifically data downloader 146 may be configured to download a portion of AI model 106 into memory 142 of target device 134a from parameter server 102.);
partitioning said AI/ML model into a plurality of sub-parts (Fig. 2 – Pudipeddi teaches downloading portions (i.e., AI model split into subparts);
forming a set of aggregation chunks (Fig. 2, ¶0055 – Pudipeddi teachesAI model dissected into layers and the downloaded portion (i.e., subpart) includes one or more layers (i.e., aggregated chunks)) and
transmitting said set of aggregation chunks to said client device (Fig. 17, network, ¶0049 – Pudipeddi teaches At step 202, a portion of an artificial intelligence (AI) model is downloaded into memory of a target device from a parameter server that stores a master copy of the AI model. For example, target device 134a of FIG. 1, specifically data downloader 146 may be configured to download a portion of AI model 106 into memory 142 of target device 134a from parameter server 102.).
Zhou discloses what Pudipeddi does not expressly disclose.
Zhou discloses:
based on (i) a first time period corresponding to an amount of time for said client device to obtain said set of aggregation chunks from said server and (ii) a second time period corresponding to an amount of time for said client device to make inferences on said set of aggregation chunks, each aggregation chunk corresponding to one or more sub-parts of said plurality of sub-parts (Section 3.2 – Partitioning a DNN Mode – Zhou teaches the building of a regression model to measure execution latency (i.e., inference costs) and then cost functions are developed for each layer and The cost function for a fused block … is then defined as the total data transfer time (i.e., time to obtain chunks) of the first layer’s input tensor, the last layer’s output tensor, and the sum of the computation time of all grouped layers.);
Pudipeddi and Zhou are analogous arts because they are from the same field of endeavor with respect to partitioning data for modeling.
Before the effective filing date, it would have been obvious to a person of ordinary skill in the art to incorporate functionality cost determinations as discussed in Zhou with AI/ML model partitioning as discussed in Pudipeddi by adding the functionality of Zhou to the system/method of Pudipeddi in order to increase optimization (Zhou, 3.2 Partitioning a DNN Model).
Yang discloses what Pudipeddi does not expressly disclose.
Yang discloses:
wherein said request includes information indicating a downlink bitrate and inference time (Fig. 3, ¶0051, ¶0052, ¶0071, ¶0076, ¶0081 – Yang teaches making a inference request that takes into account model (i.e., AI/ML models) limitations and capabilities related to speed, throughput, latency, costs (i.e., indications of downlink bitrate and inference time))
Pudipeddi, Zhou and Yang are analogous arts because they are from the same field of endeavor with respect to partitioning data for modeling.
Before the effective filing date, it would have been obvious to a person of ordinary skill in the art to incorporate request considerations as discussed in Yang with functionality cost determinations as discussed in Zhou with AI/ML model partitioning as discussed in Pudipeddi by adding the functionality of Yang to the system/method of Pudipeddi and Zhou in order to dynamically and/or statically assigning incoming inference requests for model execution (Yang, ¶0001).
Lee discloses what Pudipeddi, Zhou and Yang do not expressly disclose.
Lee discloses:
partitioning said [AI/ML model] into a plurality of sub-parts based on said downlink bitrate and inference time (¶0089 – Lee teaches that a fragment can be selected using its size, generation time and bit rate. A size of a fragment (i.e., sub-part) is based on session characteristics and known download capacity. Further, Lee teaches that the time for processing (i.e., inference time) needs to be taken into consideration for the deduced fragment size as well as bitrate attribute and the period for generating a fragment.).
Pudipeddi, Zhou, Yang and Lee are analogous arts because they are from the same field of endeavor with respect to partitioning data.
Before the effective filing date, it would have been obvious to a person of ordinary skill in the art to incorporate usage of device attribute to fragment data as discussed in Lee with request considerations as discussed in Yang with functionality cost determinations as discussed in Zhou with AI/ML model partitioning as discussed in Pudipeddi by adding the functionality of Lee to the system/method of Pudipeddi, Zhou and Yang in order to determine how to split data (Lee, ¶0089).
As to claim 3, Pudipeddi, Zhou, Yang and Lee discloses:
method of claim 1, and
Pudipeddi discloses:
wherein said first time period is based on a size of said set of aggregation chunks and a bitrate (¶0089 – Pudipeddi teaches a microbatch size (i.e., aggregated chunks) suitable for the target device is determined … by batch manager … based on the rate of communication between target device … and parameter server. Also, the batch manager may select a microbatch size based on hardware specifications for target device and then iteratively adjust it to an optimum microbatch size, for example, based on computation time for target device (i.e., inference time), size of a subportion of AI model 106 to be transmitted, and/or communication bandwidth for system 100 (i.e., time to obtain)).
As to claim 4, Pudipeddi, Zhou, Yang and Lee discloses:
method of claim 1, and
Pudipeddi discloses:
wherein said set of aggregation chunks comprises a first aggregation chuck to be transmitted first in time, and wherein said first aggregation chunk is usable for generating
(i) an inference without using other aggregation chunks of the set of aggregation chunks (¶0040 – Pudipeddi teaches AI model may be dissected into smaller portions or chunks (e.g., individual layers), and each portion or layer may be executed as efficiently as possible on the target device. After a layer is done, the next layer is executed; ¶0036 – Pudipeddi teaches that after each layer, the activations (also referred to herein as hidden activations, hidden states, or intermediate results) of that layer may be saved (e.g., on chip or off chip) for the backward pass, which may be executed in a similar layer-by-layer manner (in reverse order) on the same minibatch),
(ii) an intermediate result without using said other aggregation chunks, or
(iii) an intermediate result with previous intermediate result without using other aggregation chunks.
As to claim 6, Pudipeddi, Zhou, Yang and Lee discloses:
method of claim 1, and
Pudipeddi discloses:
wherein each sub-part corresponds to one or more neural network layers (Abstract – Pudipeddi teaches AI model may be dissected into smaller portions (e.g., layers or sub-layers), and each portion may be executed as efficiently as possible on the target device; ¶0050, ¶0051 – Pudipeddi teaches AI model 106 may have any type of deep learning architecture, for example, deep neural networks, recurrent neural networks and convolutional neural networks and a simple neural network may include several layers).
As to claim 7, Pudipeddi, Zhou, Yang and Lee discloses:
method of claim 1, and
Zhou discloses:
further comprising:
adjusting said set of aggregation chunks, at least one of updated first time period and updated second time period (Section 3.3 – Zhou teaches recalculating optimal partitions after a period of time and periodically recalculating the optimal partition points, once there is a change in the model execution graph).
As to claim 8, Pudipeddi, Zhou, Yang and Lee discloses:
method of claim 1, and
Pudipeddi discloses:
further comprising:
forming different combinations of sub-parts; and
selecting one of said combinations to form said set of aggregation chunks (¶0055 – Pudipeddi teaches the portion of AI model that is downloaded to target device may include any part of AI model up to the entirety of AI model).
As to claim 10, Pudipeddi, Zhou, Yang and Lee discloses:
method of claim 1, and
Pudipeddi discloses:
wherein each aggregation chunk includes one or more of the following:
a first identifier (ID) of that aggregation chunk (Fig. 5, L1, L22-24 of Pudipeddi),
a second ID of a preceding aggregation chunk that a current aggregation chunk is tied to,
a chunk type indicating whether a current aggregation chunk is a model entry, an intermediate chunk, or a final chunk of said AI/ML model,
a total number of aggregation chunks in said AI/ML model,
a chunk index of a current aggregation chunk,
a size of said current aggregation chunk,
an expected inference time of said current aggregation chunk on one or more target client devices,
a reference bitrate,
a reference device profile, and
a baseline model identifier.
As to claim 21, similar rejection as to claim 1.
As to claim 22, similar rejection as to claim 6.
As to claim 23, similar rejection as to claim 7.
As to claim 24, similar rejection as to claim 8.
Claims 11 and 25 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication No. 2022/0283820 A1 to Pudipeddi et al. (“Pudipeddi”) in view of Printed Publication, “Distributing Deep Neural Networks with Containerized Partitions at the Edge” to Zhou et al. (“Zhou”) in further view of U.S. Patent Application Publication No. 2021/0232399 A1 to Yang et al. (“Yang”) in further view of U.S. Patent Application Publication No. 2021/0160296 A1 to Lee et al. (“Lee”) in further view of Printed Publication, “Adaptative Inference Cost with Convolutional Neural Mixture Models” to Ruiz et al. (“Ruiz”).
As to claim 11, Pudipeddi, Zhou, Yang and Lee discloses:
method of claim 1,
Ruiz discloses what Pudipeddi, Zhou, Yang and Lee do not expressly disclose.
Ruiz discloses:
wherein said AI/ML model is a convolutional neural mixture model (Fig. 1 of Ruiz), and
wherein said AI/ML model is partitioned into a pruned convolutional neural network mixture model and (ii) one or more removed convolutional neural networks (Fig. 1 of Ruiz).
Pudipeddi, Zhou, Yang, Lee and Ruiz are analogous arts because they are from the same field of endeavor with respect to partitioning data for modeling.
Before the effective filing date, it would have been obvious to a person of ordinary skill in the art to incorporate pruning and cost analysis as discussed in Ruiz with usage of device attribute to fragment data as discussed in Lee with request considerations as discussed in Yang with functionality cost determinations as discussed in Zhou with AI/ML model partitioning as discussed in Pudipeddi by adding the functionality of Ruiz to the system/method of Pudipeddi, Zhou, Yang and Lee in order to reduce computational costs (Ruiz, Fig. 1).
As to claim 25, similar rejection as to claim 12.
Claims 15, 16, 18, 26 and 27 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication No. 2022/0283820 A1 to Pudipeddi et al. (“Pudipeddi”) in view of U.S. Patent Application Publication No. 2021/0232399 A1 to Yang et al. (“Yang”) in further view of U.S. Patent Application Publication No. 2021/0160296 A1 to Lee et al. (“Lee”).
As to claim 15, Pudipeddi discloses:
a method; implemented by a wireless transmit/receive unit (WTRU), comprising:
transmitting a request for a chunk that is part of an artificial intelligence and/or machine learning (AI/ML) model, wherein said information is transmitted from said WTRU to a server via any of wired and wireless communication (Fig. 17, network, ¶0049 – Pudipeddi teaches At step 202, a portion of an artificial intelligence (AI) model is downloaded into memory of a target device from a parameter server that stores a master copy of the AI model. For example, target device 134a of FIG. 1, specifically data downloader 146 may be configured to download a portion of AI model 106 into memory 142 of target device 134a from parameter server 102.);
receiving said chunk from said server (Fig. 17, network, ¶0049 – Pudipeddi teaches At step 202, a portion of an artificial intelligence (AI) model is downloaded into memory of a target device from a parameter server that stores a master copy of the AI model. For example, target device 134a of FIG. 1, specifically data downloader 146 may be configured to download a portion of AI model 106 into memory 142 of target device 134a from parameter server 102.);
generating a first inference or intermediate result from said chunk (¶0040 – Pudipeddi teaches AI model may be dissected into smaller portions or chunks (e.g., individual layers), and each portion or layer may be executed as efficiently as possible on the target device. After a layer is done, the next layer is executed;);
receiving, from said server, a subsequent chunk that is also part of said AI/ML model (¶0055, ¶0056 – Pudipeddi teaches target device 134a may download a next portion of AI model 106 in one or more memory buffers while executing a current portion of AI model 106. This approach may use a bit more memory and special libraries but may result in higher performance of AI model 106. In another example, target device 134a may execute the current subportion, synchronize, and then download the next subportion.); and
generating an inference result based on said first inference or intermediate result and said subsequent chunk, wherein receiving said subsequent chunk and said generating first inference or intermediate result are performed in parallel (¶0055, ¶0056 – Pudipeddi teaches target device 134a may download a next portion of AI model 106 in one or more memory buffers while executing a current portion of AI model 106. This approach may use a bit more memory and special libraries but may result in higher performance of AI model 106. In another example, target device 134a may execute the current subportion, synchronize, and then download the next subportion.).
Yang discloses what Pudipeddi does not expressly disclose.
Yang discloses:
wherein said request includes information indicating a downlink bitrate and inference time (Fig. 3, ¶0051, ¶0052, ¶0071, ¶0076, ¶0081 – Yang teaches making a inference request that takes into account model (i.e., AI/ML models) limitations and capabilities related to speed, throughput, latency, costs (i.e., indications of downlink bitrate and inference time))
Pudipeddi and Yang are analogous arts because they are from the same field of endeavor with respect to partitioning data for modeling.
Before the effective filing date, it would have been obvious to a person of ordinary skill in the art to incorporate request considerations as discussed in Yang with AI/ML model partitioning as discussed in Pudipeddi by adding the functionality of Yang to the system/method of Pudipeddi in order to dynamically and/or statically assigning incoming inference requests for model execution (Yang, ¶0001).
Lee discloses what Pudipeddi and Yang do not expressly disclose.
Lee discloses:
partitioning said [AI/ML model] into a plurality of sub-parts based on said downlink bitrate and inference time (¶0089 – Lee teaches that a fragment can be selected using its size, generation time and bit rate. A size of a fragment (i.e., sub-part) is based on session characteristics and known download capacity. Further, Lee teaches that the time for processing (i.e., inference time) needs to be taken into consideration for the deduced fragment size as well as bitrate attribute and the period for generating a fragment.).
Pudipeddi, Yang and Lee are analogous arts because they are from the same field of endeavor with respect to partitioning data.
Before the effective filing date, it would have been obvious to a person of ordinary skill in the art to incorporate usage of device attribute to fragment data as discussed in Lee request considerations as discussed in Yang with AI/ML model partitioning as discussed in Pudipeddi by adding the functionality of Lee to the system/method of Pudipeddi and Yang in order to determine how to split data (Lee, ¶0089).
As to claim 16, Pudipeddi, Yang and Lee discloses:
method of claim 15, and
Pudipeddi discloses:
wherein generation of said first inference or intermediate result from said chunk starts as soon as said chunk is received (¶0040 – Pudipeddi teaches AI model may be dissected into smaller portions or chunks (e.g., individual layers), and each portion or layer may be executed as efficiently as possible on the target device. After a layer is done, the next layer is executed;).
As to claim 18, Pudipeddi, Yang and Lee discloses:
method of claim 15, and
Pudipeddi discloses:
further comprising:
reevaluating at least one of (i) a first time period corresponding to an amount of time for said WTRU to obtain said chunk and (ii) a second time period corresponding to an amount of time for said WTRU to make inferences on said chunk (¶0056 – Pudipeddi teaches a microbatch size may be initially selected for target device 134a based on its hardware specifications, and then the microbatch size may be adjusted in an iterative process as needed to adequately hide the communication latency. In embodiments, an optimal microbatch size may be a tradeoff between the memory required and the percentage of communication overhead that may be hidden. As more communication overhead is hidden, more memory may be required for computation. Thus, a microbatch size may be large enough to fully utilize a layer's execution in a target device, but is small enough to fit into the memory of that target device.); and
requesting said server to adjust how chunks are generated (¶0076 – Pudipeddi teaches For example, output data manager 124 of FIG. 1 may receive gradients from target devices 134a-134k. Gradients are the adjustments calculated by backpropagating the error in predictions through an AI model. Thus, gradients are values representing the difference between where the weights of the model are versus where the weights should be. The gradients may be placed in data structures, such as matrices. In an example embodiment, the gradients may be received after execution of every microbatch).
As to claim 26, similar rejection as to claim 15.
Claim 17 is rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication No. 2022/0283820 A1 to Pudipeddi et al. (“Pudipeddi”) in view of U.S. Patent Application Publication No. 2021/0232399 A1 to Yang et al. (“Yang”) in further view of U.S. Patent Application Publication No. 2021/0160296 A1 to Lee et al. (“Lee”) in view of U.S. Patent Application Publication No. 2022/0215251 A1 to Liu et al. (“Liu”).
As to claim 17, Pudipeddi, Yang and Lee discloses:
method of claim 15,
Liu discloses what Pudipeddi, Yang and Lee does not expressly disclose.
Liu discloses:
further comprising
deleting said chunk after said first inference or intermediate result is generated (¶0096 – Liu teaches deleting input from memory after it is used).
Pudipeddi, Yang, Lee and Liu are analogous arts because they are from the same field of endeavor with respect to partitioning data for modeling.
Before the effective filing date, it would have been obvious to a person of ordinary skill in the art to incorporate removing unneeded data as discussed in Liu with usage of device attribute to fragment data as discussed in Lee with request considerations as discussed in Yang with AI/ML model partitioning as discussed in Pudipeddi by adding the functionality of Liu to the system/method of Pudipeddi and Yang in order to demonstrate to removal of data that is no longer needed and provide space for needed information (Ruiz, ¶0035).
Claim 27 is rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication No. 2022/0283820 A1 to Pudipeddi et al. (“Pudipeddi”) in view of U.S. Patent Application Publication No. 2021/0232399 A1 to Yang et al. (“Yang”) in further view of U.S. Patent Application Publication No. 2021/0160296 A1 to Lee et al. (“Lee”) in further view of U.S. Patent Application Publication No. 2014/0108495 A1 to Benno et al. (“Benno”).
As to claim 27, Pudipeddi, Yang and Lee discloses:
WTRU of claim 26, and
Lee discloses:
reevaluate at least one of (i) a first time period corresponding to an amount of time for said WTRU to obtain said chunk and (ii) a second time period corresponding to an amount of time for said WTRU to make inferences on said chunk (¶0089 – Lee teaches that a fragment can be selected using its size, generation time and bit rate. A size of a fragment (i.e., sub-part) is based on session characteristics and known download capacity. Further, Lee teaches that the time for processing (i.e., inference time) needs to be taken into consideration for the deduced fragment size as well as bitrate attribute and the period for generating a fragment.); and
Benno discloses what Pudipeddi, Yang and Lee do not expressly disclose.
Benno and Lee discloses:
request said server to adjust how chunks are generated (Fig. 3, ¶0023 – Benno shows steps to update how chunks are generated and eventually downloaded).
Pudipeddi, Yang, Lee and Benno are analogous arts because they are from the same field of endeavor with respect to partitioning data for modeling.
Before the effective filing date, it would have been obvious to a person of ordinary skill in the art to incorporate repartitioning as discussed in Benno with usage of device attribute to fragment data as discussed in Lee with request considerations as discussed in Yang with AI/ML model partitioning as discussed in Pudipeddi by adding the functionality of Benno to the system/method of Pudipeddi, Yang and Lee in order to adjust quality of data received by the client (Benno, ¶0008).
Claim 28 is rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication No. 2022/0283820 A1 to Pudipeddi et al. (“Pudipeddi”) in view of U.S. Patent Application Publication No. 2021/0232399 A1 to Yang et al. (“Yang”) in further view of U.S. Patent Application Publication No. 2021/0160296 A1 to Lee et al. (“Lee”) in further view of Printed Publication, “Adaptative Inference Cost with Convolutional Neural Mixture Models” to Ruiz et al. (“Ruiz”).
As to claim 28, Pudipeddi, Yang and Lee discloses:
WTRU of claim 26,
Ruiz discloses what Pudipeddi and Yang does not expressly disclose.
Ruiz discloses:
wherein said AI/ML model is a convolutional neural mixture model, and wherein said AI/ML model is partitioned into a pruned convolutional neural network mixture model and (ii) one or more removed convolutional neural networks (Fig. 1 of Ruiz).
Pudipeddi, Yang, Lee and Ruiz are analogous arts because they are from the same field of endeavor with respect to partitioning data for modeling.
Before the effective filing date, it would have been obvious to a person of ordinary skill in the art to incorporate pruning and cost analysis as discussed in Ruiz usage of device attribute to fragment data as discussed in Lee with request considerations as discussed in Yang with AI/ML model partitioning as discussed in Pudipeddi by adding the functionality of Ruiz to the system/method of Pudipeddi, Yang and Lee in order to reduce computational costs (Ruiz, Fig. 1).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to TAYLOR A ELFERVIG whose telephone number is (571)270-5687. The examiner can normally be reached Monday (10:00 AM CST) - Friday (4:00 PM CST).
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, Oscar Louie can be reached at (571) 270-1684. 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.
/TAYLOR A ELFERVIG/Primary Examiner, Art Unit 2445
Prosecution Insights