DETAILED ACTION
This Action is responsive to Claims filed 05/22/2026.
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 .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 05/22/2026 has been entered.
Status of the Claims
Claims 1, 6-7, 11, 19-20, 24, and 28 have been amended. Claims 1-4, 6-8, 10-14, 16-21, 23-26, and 28-29 are currently pending.
Response to Arguments
Applicant’s arguments, see Pages 10-13, filed 05/22/2026, regarding the 35 U.S.C. 103 Rejection(s) of claims 1-4, 6-8, 10-14, 16-21, 23-26, and 28-29 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Claim Rejections - 35 USC § 103
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claim(s) 1-4, 6, 8, 10-14, 16-19, 21, 23-26, and 28-29 is/are rejected under 35 U.S.C. 103 as being unpatentable over Liu Shengli et al. (CN 111401552 A), hereinafter Liu; further in view of Du et al. (US 11,017,322 B1), hereinafter Du; Module MULTIMEDIA-SYSTEM-CONTROL (H.245:01/2005), hereinafter H.245; Prakash et al (US 2019/0138934 Al), hereinafter Prakash; Schlezinger et al. (Federated Learning with Quantization Constraints, 2020), hereinafter Schlezinger; and Zhang et al. (Staleness-aware Async-SGD for Distributed Deep Learning, 2016), hereinafter Zhang.
In regards to claim 1: The present invention claims: “A method for wireless communication at a user equipment (UE), comprising:” Liu teaches a federal learning method with a plurality of terminals (mapping user equipment to terminals), with wireless communication (Abstract).
“receiving first information for updating parameters of a machine learning model;” Liu teaches a terminal receiving an adjusted batch size from an edge server, which is an adjustment to the model (mapping first information for updating parameters to receiving an adjusted batch size) (Page 3, starting at “the edge server adjusts the batch size…”).
“transmitting, via the channel, compressed gradient data that is generated based at least in part on the gradient data output by the machine learning model and the quantization level,” Liu teaches the terminal transmitting the compressed gradient information to the edge server according to the compression rate (Page 3, “the terminal performs model learning according to the received batch size, compresses gradient information obtained by the model learning according to the received extraction compression rate and outputs the gradient information to the edge server;”).
“wherein the gradient data output by the machine learning model is based at least in part on updating the machine learning model using the first information;” Liu teaches the model updating with the received batch size prior to outputting the data (Page 3, starting at “the terminal performs model learning according to…”).
While Liu does teach a terminal receiving an adjusted gradient compression ratio from an edge server (Page 3), it fails to explicitly teach “receiving an indication of a quantization level of the set of quantization levels for gradient data output by the machine learning model, the quantization level…comprising a quantity of bits used to transmit the compressed gradient data…” However, Du teaches “The DP intrinsic quantization includes mapping coordinates in the updated gradient vector to a set of discrete finite values. Furthermore, to provide a controllable DP protection protocol, the DP intrinsic quantization is designed to be adaptive to the local subsampling size of data records at each client in each round of FL training. In some embodiments, the DP intrinsic quantization includes: determining a global quantization hyper-parameter for the client 310; determining a delta based on the global quantization hyper-parameter and a quantity of the plurality of data records; (data records received from server, used in determining quantization level, mapping to “an indication”) and performing the quantization based on a space determined based on the delta. This delta here may define a quantization space with a plurality of quantization steps. For example, if a quantization range is from -5 to 5 and the delta is 2, the quantization space may be defined with the following quantization steps: (-5, -3), (-3, -1), (-1 , 1), (1, 3), and (3, 5). When performing quantization, a floating number (e.g., 32 bits) within a quantization step may be mapped to a corresponding integer (e.g., 5 quantization steps may correspond to 5 integers, which may be represented by as few as 3 bit). (quantity of bits) In some embodiments, the performing of the quantization includes: determining a number of levels of the quantization based on the delta; and performing the quantization based on the range and the number of levels. (quantization based on indication and plurality of bits)” (Column 9, Lines 19-45).
Du highlights the deficiencies of the art at the time of their filing, both in regards to privacy and data compression in federated learning systems (Background). It would have been obvious to one of ordinary art at the time of the applicant’s filing to combine the federated learning system of Liu with the methods of Du to increase the privacy and data compression of transmitting quantized data to and from the clients and server.
While Liu teaches the terminals uploading information about themselves prior to model adjustment. This information includes computing power (Page 8, starting at “701. initialization, each terminal needs to upload related information…”); and Du also teaches quantization levels and the server performing client selection for determining which clients will partake in a current round of federated learning (Column 7, Lines 8-20), the combination of Liu and DU fails to explicitly teach the transfer of supported quantization levels in their respective capability messages as recited in “transmitting, to a server, a capability message indicating a set of quantization levels supported by the UE, wherein quantization levels in the set of quantization levels are supported by the UE for communication to the server;” However, H.245 (a protocol utilized since 1996 or earlier, per Lindbergh et al. (not used in the rejection herein, referenced merely to indicate the well-known implementation of such capability information (Lindbergh Pages 1 and 6, at least))), shares compression/quantization capabilities of an edge device with a server during a communication handshake (Page 4 shows the DataApplicationCapability exchange, Page 16 indicates compression capabilities in DataApplicationCapability and DataProtocolCapability).
H.245 shows a protocol for capability exchange between an edge device and central controller conveying compression capabilities of the edge device would have been well-known in the field of federated communication networks well before the Applicant’s filing date. It would have obvious to one skilled in the art, implementing a similar communication system to transferring federated learning model information, to use such well-known protocols.
While a combination of Liu, Du, and H.245 teaches gradient quantization levels and the terminals sending initialization data, including the state of the wireless channel, before federal learning begins (Page 8, starting at “701. initialization, each terminal needs to upload related information…”), Liu fails to explicitly teach “based at least in part on a bandwidth of a channel for transmitting compressed gradient data and a link budget associated with the UE…” However, Prakash teaches a distributed or federal learning system with heterogenous compute nodes send operational parameters to the master node ([0095]). Prakash paragraphs [0048]-[0050] illustrate the information that could be included in the operational parameters sent to the master node, including network capabilities, channel states, and network traffic measurements or capabilities.
Prakash addresses the difficulty heterogenous federal learning systems can have with edge nodes and wireless links of various strengths and computational capacities and the need to base the edge node’s work on the broader model on balancing the load for each node’s capabilities (Abstract, Background, [0028]-[0029]). It would have been obvious to one of ordinary skill in art before the applicant’s filing date to combine a system such as the one in Liu, with the efficiency methods of Du, and communicate terminal capability information during initialization like the operational parameters described in Prakash and/or with H.245 to properly balance the computational load of the federated learning system and increase the overall efficiency of model convergence.
While Du teaches the quantization levels being higher or lower based on local and global deltas (Column 9, Lines 46-60), the combination of Liu, Du, H.245, and Prakash fails to explicitly teach “wherein a higher quantization level corresponds to a higher data rate of the channel;” However, Shlezinger teaches “In particular, we identify the unique characteristics associated with conveying trained models over rate-constrained channels, and characterize a suitable quantization scheme for such setups.” (Abstract) and “This can be handled by letting the encoder use a lattice with a basic cell P0 of increased volume if the resulting uk exceeds the bit limitation Rk, decreasing the number of bits at the cost of a more coarse quantization.” (Page 8853, Left Column).
Shlezinger teaches “We show that combining universal vector quantization methods with FL yields a decentralized training system, which is both efficient and feasible. We also derive theoretical performance guarantees of the system.” (Abstract). It would have been obvious to one of ordinary skill in the art at the time of the Applicant’s filing to combine a system such as the one in Liu, with the efficiency methods of Du, communicate terminal capability information like described in Prakash and H.245, and account for rate-constrained channels like Shlezinger teaches to increase the efficiency and performance of a federated learning system.
While the combination of Liu, Du, H.245, Prakash, and Shlezinger fails to explicitly teach “and transmitting, to the server and via the channel, an indication of a time at which the gradient data is output by the machine learning model.” Zhang, in a similar field of endeavor, teaches the use of a timestamp value, incremented when a weight update occurs, assigned to a gradient and the weight used to calculate it (Section 2.1, Page 2) as well as in “To advance the weights’ timestamp θfrom i to i + 1, each learner l compute a gradient…using a mini-batch size of μ and sends it to the parameter server.” (Section 2.2, Page 2).
Zhang teaches “However, permitting this asynchronous behavior inevitably adds “staleness” to the system wherein some of the workers compute gradients using model parameters that may be several gradient steps behind the most updated set of model parameters.” (Introduction) and the calculation of a learning rate based on the timestamp and staleness in Section 2.1. Similar to the other information sent back and forth between clients and servers in a federated learning system, a person of ordinary skill in the art at the time of the Applicant’s filing would have been aware of sending gradient or weight timestamps in order to mitigate staleness.
In regards to claim 2: The present invention claims: “The method of claim 1, further comprising: compressing the gradient data output by the machine learning model based at least in part on the quantization level, wherein the transmitting of the compressed gradient data is based at least in part on the compressing of the gradient data.” While Liu teaches that the data is compressed based on a received gradient compression ratio before the compressed data is transmitted (Page 3, starting at “the terminal performs model learning according to the received batch size, compresses gradient information obtained by the model learning according to the received extraction compression rate and outputs the gradient information to the edge server;”) it fails to explicitly teach “the quantization level”, However, see the above rejection for claim 1 how Du teaches the determination of a quantization level and its use on updated model parameters before it is transmitted (Fig. 5, number 530) and how to combination of Liu and Du would have been obvious to one of ordinary skill in the art at the time of the applicant’s filing.
In regards to claim 3: The present invention claims: “The method of claim 1, further comprising: receiving second information for updating the parameters of the machine learning model based at least in part on the transmitting of the compressed gradient data;” Liu teaches the method for federated learning iteratively adjusting (See Page 3, “and (c) repeating the step (a) and the step (b) until the current learning time delay is equal to the specified learning time delay…”, where step (b) adjusts the batch size and compression rate), sending, and receiving the batch size, gradient compression ratio, and gradient data output until time delay threshold is met (Page 3, starting at “(a) calculating the current learning time delay according to the current batch size…”).
“receiving a second indication of a second quantization level for second gradient data output by the machine learning model, the second quantization level based at least in part on a duration associated with the communicating of the compressed gradient data;” See above how Du teaches a similarly used quantization level. Liu teaches the method for federated learning iteratively adjusting (See Page 3, “and (c) repeating the step (a) and the step (b) until the current learning time delay is equal to the specified learning time delay…”, where step (b) adjusts the batch size and compression rate), sending, and receiving the batch size, gradient compression ratio, and gradient data output until time delay threshold is met (Page 3, starting at “(a) calculating the current learning time delay according to the current batch size…”). See also where Liu teaches the calculation of the time delay including the time it takes to send and receive the data (Page 3, starting at “In the step (a), the calculation process of the current learning time delay…”).
It would have obvious for one of ordinary skill in the art at the time of the applicant’s filing to further combine Liu and Du and account for a transmittal time delay, in the event the quantization hyper-parameter or delta needed adjusting on a client to further optimize the compression of data for transmittal.
“and transmitting second compressed gradient data that is generated based at least in part on the second gradient data output by the machine learning model and the second quantization level, wherein the second gradient data output by the machine learning model is based at least in part on updating the machine learning model using the second information.” Liu teaches the method for federated learning iteratively adjusting, sending, and receiving the batch size, gradient compression ratio, and gradient data output until time delay threshold is met (Page 3, starting at “(a) calculating the current learning time delay according to the current batch size…”). Du also teaches quantization levels and iterating the training process multiple times (Column 8, Lines 27-35).
In regards to claim 4: The present invention claims: “The method of claim 3, wherein: the quantization level is associated with a set of UEs that includes the UE; and the second quantization level is specific to the UE.” While Liu teaches the gradient compression ratio being generally adjusted before being distributed to the terminals, where further adjusts in later iterations will be based on updates the terminals perform themselves (Page 3), it fails to explicitly teach “quantization level” However, see above where Du teaches quantization levels. Du also teaches global quantization hyper-parameters as well as a quantization delta for specific terminal devices (Column 2, Lines 54-65).
In regards to claim 6: The present invention claims: “The method of claim 1, further comprising: and transmitting an indication of the quantization level used to compress the gradient data output by the machine learning model.” See the arguments for claim 3 for how a combination of Liu and Du teach how the batch size, quantization levels, gradient data, and time delay are calculated iteratively until a threshold in the time delay is met (Page 3, specifically steps (a), (b), and (c), where time delay is calculated, parameters are adjusted and resent, and both aforementioned steps are iterated until a threshold is reached (mapping iteration of these steps to subsequent indications of parameters/times).
In regards to claim 8: The present invention claims: “The method of claim 1, further comprising: receiving a set of quantization levels that includes the quantization level, wherein the indication of the quantization level identifies the quantization level from the set of quantization levels that is for the UE.” Liu teaches a set of gradient values being used to arrive at an average gradient value. The calculations include indices for individual terminals amid the set of terminals (Page 4, starting at “In one possible implementation, the gradient average…”). A combination with Du would also include the global quantization hyper-parameters for each terminal device. See above how the use of H.245 or similar protocol would communicate the compression/quantization capabilities of an edge device.
In regards to claim 10: The present invention claims: “The method of claim 1, wherein the machine learning model comprises a federated learning model associated with a set of UEs including the UE, and wherein each UE of the set of UEs is associated with a unique dataset of the machine learning model.” Liu teaches a federated learning model with a set of terminals with local training data to each terminal (Abstract, Page 3, starting at “in a first aspect, a federated learning method based on batch size…”).
In regards to claim 11: The present invention claims: “A method for wireless communication at a server, comprising: receiving, from a user equipment (UE) of a set of UEs, a capability message indicating a set of quantization levels supported by the UE for gradient data output by a machine learning model implemented by the UE; transmitting, to the UE, first information for updating parameters of the machine learning model and an indication of quantization level of the set of quantization levels for the gradient data output by the machine learning model; and receiving, from the UE, compressed gradient data based at least in part on the transmitting of the first information and the indication of the quantization level; and receiving, from the UE and via the channel, an indication of a time which the gradient data is output by the machine learning model.” The only substantial difference between claim 11 and claims 1 and 5 is the recitation of “A method for wireless communication at a server, comprising: determining, for a user equipment (UE) of a set of UEs, a quantization level for gradient data output by a machine learning model implemented by the UE;” However, Liu teaches an edge server that determines a batch size and gradient compression ratio as part of their system (Page 3, starting at “in a first aspect, a federated learning method based on batch size…”).
In regards to claim 12: See the above arguments for how a combination of Liu and Du teach calculating a time delay (duration), and iteratively sends adjusted batch sizes and quantization levels, receives output data (gradient data) until a predetermined time delay is met (satisfying a threshold duration).
In regards to claim 13: See the above arguments on claim 3 for how Liu teaches the calculation of a time delay.
In regards to claim 14: See the above arguments on claim 4 for how a combination of Liu and Du teaches a general gradient compression ratio first then subsequent iterations being specific to the terminals, and how that would apply to sending and receiving quantization levels.
In regards to claim 16: See the above arguments for how a combination of Liu and Du teach an edge server and terminals communicating and sending batch sizes (first information), quantization levels, and compressed gradient data from the terminals.
In regards to claim 17: See the above arguments for how a combination of Liu and Du teaches iteratively sending batch sizes and quantization levels to the terminals. Information on Liu teaching an average or mean of the terminal’s data can also be found in the same locations as the above references and Page 9.
In regards to claim 18: See the above arguments for how Liu teaches a federated learning system, where each terminal calculates gradient data and sends it back to the edge server to achieve model convergence.
In regards to claim 19: See the above arguments for how a combination of Liu and Du teach the edge server and terminals iteratively communicating quantization levels to compress the gradient data of the terminals.
In regards to claim 21: See the above arguments for claim 8 for how Liu teaches an average gradient value and the use of indices to indicate specific terminals.
In regards to claim 23: See the above arguments for claim 10 for how Liu teaches a federated learning model with a set of terminals with local training data to each terminal.
In regards to claims 24-26 and 28-29: The only significant difference between the limitations of claims 24-26 and 28-29 and the preceding method claims are that claims 24-26 and 28-29 recite an “An apparatus for wireless communication…” at a user equipment and server with “a processor; memory coupled with the processor; and instructions stored in the memory and executable by the processor” However, given the apparatus of claims 24-29 implement the method of claims 1-23, both sets of claims are similarly rejected.
Claim(s) 7 and 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Liu, Du, H.245, Prakash, Schlezinger, and Zhang as applied to claims 1 and 11 above, and further in view of Brendan McMahan et al. (US 2019/0227980 Al), hereinafter McMahan.
In regards to claim 7: The present invention claims: “The method of claim 6, wherein the transmitting of the indication of the quantization level comprises: transmitting a set of quantization levels for the gradient data output by the machine learning model,” See the above arguments for how a combination of Liu, Du, and H.245 teaches sending and receiving batch sizes and quantization levels or information pertaining to quantization levels, and gradient data iteratively (mapping iteration of time delay and adjustments to parameters to subsequent indication(s)) and how a combination of Liu, Du, and H.245 teaches the communication of terminal device capability and their selection by the server.
Liu, Du, and Prakash fail to explicitly teach “…each quantization level of the set of quantization levels associated with a dimensional parameter of a set of dimensional parameters associated with the gradient data output by the machine learning model.” However, McMahan teaches the client computing devices in a federated learning system determining the local updates to their models by performing some number of mini-batch stochastic gradient descent steps ([0025]). Determining the local update can further include clipping the local model. This clipping can be compared to the server model, or a layer or per-layer clipping (mapping layer/per-layer clipping in model as dimensional parameter, mapping the inclusion of both gradient descent steps and clipping as association with the output) ([0027-[0030]).
McMahan highlights the need to increase accuracy and security in increasingly complex machine learning systems (Background). Given the variable nature of size and complexity of machine learning models, it would have been obvious to one of ordinary skill in art before the applicant’s filing date to combine a system such as the one in Liu, with the efficiency methods of Du, and the channel capabilities of Prakash and include parameters or data indicating the dimensional complexity of the terminal or client systems to increase the accuracy of the overall machine learning model.
In regards to claim 20: See the above arguments for claim 7, see also that the server receives said updates.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GRIFFIN T BEAN whose telephone number is (703)756-1473. The examiner can normally be reached M - F 7:30 - 4:30.
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, Li Zhen can be reached at (571) 272-3768. 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.
/GRIFFIN TANNER BEAN/Examiner, Art Unit 2121
/Li B. Zhen/Supervisory Patent Examiner, Art Unit 2121
Prosecution Insights