Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claim 9 rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claim 9 does not disclose what C is supposed to represent in the disclosed formula. Although ci is described as the individual training cost per sample for the i-th FL ML model, C is not described as anything, despite being present in the formula.
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.
Claims 1, 11, 13, 14, 20 are rejected under 35 U.S.C. 103 as being unpatentable over Wang (US 11521090 B2) in view of You (Ranking and Tuning Pre-trained Models: A New Paradigm for Exploiting Model Hubs).
Regarding claim 1, Wang discloses “receive, at an edge device, one or more federated learning, FL, machine learning, ML, models and capabilities information associated with one or more FL ML models;” (See [Page 13, Col 6, Lines 38-64]; Wang discloses a method of distributed machine learning that uses a system of edge nodes (devices) that receive information associated with ML models)
“compute utility… based on a quality of available training samples at the edge device for training the corresponding FL ML models and the capabilities information;” (See [Page 14, Col 7, Lines 15 - 18], [Page 14, Col 7, Lines 21 - 25]; Wang discloses computing the utility of edge nodes and their datasets and then determining how many training iterations to perform for an edge node)
“perform training… until a remaining available cost budget of a total available cost budget of the edge device expires” (See [Page 14, Col 7, Lines 33 - 48]; Wang discloses training edge nodes with ML models and decides to train or exclude training based on a total budget. If the total budget expires, then it could be concluded that the remaining nodes that have yet to be trained will be excluded from training).
Wang fails to explicitly disclose, “compute utility of the one or more FL ML models”
“rank the one or more FL ML models in a descending order of utility based on the computed utility of the one or more FL ML models”
“perform training of the corresponding FL ML models in the descending order of utility”.
You teaches “compute utility of the one or more FL ML models” (See [Page 3, Section 1, Paragraph 5]; You discloses the LogME method, which is used for determining the transferability of a PTM, which is a specific application of utility in a ML model).
“rank the one or more FL ML models in a descending order of utility based on the computed utility of the one or more FL ML models” (See [Page 3, Section 1, Paragraph 5]; You discloses ranking PTMs (pre-trained models) in descending order by using a LogME algorithm on the PTMs, which computes the transferability of a PTM, which is a specific application of utility in a ML model.)
“perform training of the corresponding FL ML models in the descending order of utility” (See [Page 3, Section 1, Paragraph 5]; You discloses tuning (training) models based on descending order of utility based on the top-K ranked PTMs (pre-trained models))
Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention having Wang and You before them to modify Wang to compute the utility for models, rank models by descending order of utility, and train models by descending order of utility. One would be motivated to compute the utility for models and rank the models by descending order of unity so that a user could select the most optimal model that fits their requirements, see e.g., [Page 2, Section 1, Paragraph 4], [Page 3, Section 1, Paragraph 5], where You describes that users naively pick the most popular pre-trained model from online collections instead of the most optimal model, and determining a utility for each model and ranking by utility can show what the most optimal models are. One would also be motivated to train models by descending order of utility to prioritize training the most optimal models, see e.g., [Page 17, Section 5, Paragraph 2], where You describes that there are budget constraints on power and memory consumption when training multiple models, so prioritizing the most optimal models to train first can help with managing budget constraints.
Regarding claim 11, Wang fails to explicitly disclose, “wherein the total available cost budget of the edge device comprises any one or more cost budgets from the group of: an energy consumption budget; a time budget; a communication budget; a processing budget; any other cost budget for tracking the cost of training one or more FL ML model on the edge device; one or more combinations of cost budgets from the group; or any other type of overall cost budget of the edge device”.
You teaches “wherein the total available cost budget of the edge device comprises any one or more cost budgets from the group of: an energy consumption budget; a time budget; a communication budget; a processing budget; any other cost budget for tracking the cost of training one or more FL ML model on the edge device; one or more combinations of cost budgets from the group; or any other type of overall cost budget of the edge device” (See [Page 17, Section 5, Paragraph 2]; You discloses that the budget takes memory and power consumption into consideration when deploying neural networks, and states that the best ranked pre-trained model is not always the best choice due to these limitations).
Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention having Wang and You before them to modify Wang to consider memory and power consumption the budget for training a ML model. One would be motivated to do so to determine if the models that are currently being trained have a combined budget that is less than or equal to the total available cost budget, see e.g., [Page 17, Section 5, Paragraph 2], where You describes that using the best ranked model may not always be the most economical choice due to its memory and power consumption costs, so choosing other models that have similar rankings may result in a lower combined budget cost.
Regarding claim 13, Wang discloses “wherein the FL ML model comprises any ML model from the group of: neural network; transformer-based ML model; encoder-decoder ML model; or any other ML model capable of being trained based on available training samples stored on the edge device.” (See [Page 15, Col 10, Lines 34 - 44]; Wang discloses using a distributed machine learning model, which is a ML model capable of being trained based on available training samples stored on the edge device)
Regarding claims 14 and 20, these claims are similar in scope to claim 1.
Claim Rejections - 35 USC § 103
Claims 2, 3, 9, 15, 16 are rejected under 35 U.S.C. 103 as being unpatentable over Wang (US 11521090 B2) in view of You (Ranking and Tuning Pre-trained Models: A New Paradigm for Exploiting Model Hubs), and further in view of Das (US 20200151573 A1).
Regarding claim 2, Wang-You fails to explicitly disclose, “wherein the available training samples at the edge device for training the corresponding FL ML model further comprises at least one of a set of high-loss training samples and a set of low-loss training samples associated with training the corresponding FL ML model”.
Das teaches “wherein the available training samples at the edge device for training the corresponding FL ML model further comprises at least one of a set of high-loss training samples and a set of low-loss training samples associated with training the corresponding FL ML model” (See [0034]; Das discloses grouping a set of high-loss samples together in a first set, which grouping a set of low-loss samples together in a second set).
Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention having Wang-You and Das before them to modify Wang-You to include a set of high-loss and low-loss training samples for training a model. One would be motivated to use high-loss or low-loss sets depending on which model is being used in order to better fit the model’s characteristics, see e.g., [0013], where Das describes using high-loss sets for higher precision operands and using lower-loss sets for lower precision operands.
Regarding claim 3, Wang discloses “assign the one or more FL ML models an individual cost budget of the total available cost budget for the edge device for training the corresponding FL ML models based, at least in part, on the capabilities information associated with the corresponding FL ML models;” (See [Page 14, Col 7, Lines 21 - 32], [Page 14, Col 7, Lines 33 - 38]; Wang discloses assigning each edge node in a model an individual cost, and then decides whether to exclude the node from training or not based on the total budget)
“wherein the training of the corresponding FL ML models further comprises… training an FL ML model on the selected training samples is less than a remaining available cost budget whilst the total utility gain for the edge device is maximized.” (See [Page 11, Col 2, Lines 20 - 27], [Page 15, Col 9, Lines 1 - 4]; Wang discloses reducing the number of edge nodes being used for a model until their combined cost budget is lower than the target cost budget (remaining available cost budget) of the model requester node and also discloses obtaining a cost for each edge node, however, Wang does not disclose using a proportion of high-loss and low-loss training samples on the model).
Wang-You fails to explicitly disclose, “wherein the training of the corresponding FL ML models further comprises; select for the one or more FL ML models a proportion of high-loss training samples and low-loss training samples on the edge device such that the aggregate individual cost budget for training an FL ML model”.
Das teaches “wherein the training of the corresponding FL ML models further comprises; select for the one or more FL ML models a proportion of high-loss training samples and low-loss training samples on the edge device such that the aggregate individual cost budget for training an FL ML model” (See [0042]; Das discloses determining whether to alter the proportion of high loss or low loss training samples for model parameters depending on the desired outcome).
Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention having Wang-You and Das before them to modify Wang-You to determine a proportion of high-loss and low-loss training samples for each model to obtain a total cost budget that is less than or equal to the remaining available cost budget. One would be motivated to adjust the proportions for each model to ensure that the total cost budget doesn’t exceed the remaining available cost budget, see e.g., [0042], where Das decides to alter the proportion of high-loss and low-loss subsets if the samples become more concentrated at the high-loss or low-loss end to ensure a more even distribution of high-loss and low-loss sets, and e.g., [Page 11, Col 2, Lines 20 – 27], where Wang adjusts a number of edge nodes being used to obtain a total cost budget that is less than or equal to the remaining available cost budget.
Regarding claim 9, Wang discloses “wherein the performing of the training of the corresponding FL ML models further comprises: select a set of training samples for each of a number N of FL ML models whilst ensuring the following condition is met:
c1(SL1 + a1SH1) + c2(SL2 + a2SH2) + ... + cN(SLN + aNSHN) ≤ C
where, ci is the individual training cost per sample for the i-th FL ML model for 1 ≤ i ≤ N” (See [Page 11, Col 2, Lines 20 - 27], [Page 15, Col 9, Lines 1 - 4]; Wang discloses reducing the number of edge nodes being used for a model until their combined cost budget is lower than the target cost budget (remaining available cost budget) of the model requester node and also discloses obtaining a cost for each edge node)
Wang-You fails to explicitly disclose, “SLi is the number of selected low-loss training samples in the selected set of training samples for the i-th selected FL ML model, SHi is the number of selected high-loss training samples in the selected set of training samples for the i-th selected FL ML model 104i, and αi is a binary value in which αi = 0 when ULi>UALLi otherwise αi = 1”.
Das teaches “SLi is the number of selected low-loss training samples in the selected set of training samples for the i-th selected FL ML model, SHi is the number of selected high-loss training samples in the selected set of training samples for the i-th selected FL ML model 104i, and αi is a binary value in which αi = 0 when ULi>UALLi otherwise αi = 1” (See [0042]; Das discloses determining whether to alter the proportion of high-loss or low-loss training samples for model parameters depending on the desired outcome).
Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention having Wang-You and Das before them to modify Wang-You to determine and alter the proportion of high-loss or low-loss training samples for model parameters. One would be motivated to adjust the proportions for each model to ensure that the total cost budget doesn’t exceed the remaining available cost budget, see e.g., [0042], where Das decides to alter the proportion of high-loss and low-loss subsets if the samples become more concentrated at the high-loss or low-loss end to ensure a more even distribution of high-loss and low-loss sets, and e.g., [Page 11, Col 2, Lines 20 – 27], where Wang adjusts a number of edge nodes being used to obtain a total cost budget that is less than or equal to the remaining available cost budget.
Regarding claim 15, this claim is similar in scope to claim 2.
Regarding claim 16, this claim is similar in scope to claim 3.
Claim Rejections - 35 USC § 103
Claims 6, 18 are rejected under 35 U.S.C. 103 as being unpatentable over Wang (US 11521090 B2) in view of You (Ranking and Tuning Pre-trained Models: A New Paradigm for Exploiting Model Hubs), and further in view of KT (US 20230359609 A1).
Regarding claim 6, Wang discloses “the performance of the computing the utility” (See [Page 14, Col 7, Lines 21 - 32]; Wang discloses computing the utility)
Wang fails to explicitly disclose, “the ranking the FL ML models”.
You teaches “the ranking the FL ML models” (See [Page 3, Section 1, Paragraph 5]; You discloses ranking PTMs (pre-trained models)).
You fails to explicitly disclose, “repeat, in each time period… performing training for any untrained FL ML models until the remaining available cost budget of the total available cost budget for the each time period expires”.
KT teaches “repeat, in each time period… performing training for any untrained FL ML models until the remaining available cost budget of the total available cost budget for the each time period expires” (See [0065], [0068]; KT discloses repeatedly performing training on a model (time periods) until a budget is reached for an epoch (time period)).
Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention having Wang, You, and KT before them to modify Wang to rank the models, perform training on models until a budget is reached during a time period, as well as performing utility computation, ranking, and training repeatedly for each time period. One would be motivated to rank the models so that a user could select the most optimal model that fits their requirements, see e.g., You[Page 2, Section 1, Paragraph 4], You[Page 3, Section 1, Paragraph 5], where You describes that users naively pick the most popular pre-trained model from online collections instead of the most optimal model. One would be motivated to perform training on models until a budget is reached or almost reached during a time period so that the training may end before the remaining available cost budget is fully reached, see e.g., KT[0068], where KT describes ending training of a model before its budget is reached if its target accuracy is reached during training, or ending training when the budget is reached to cease training so that the budget is not surpassed. One would also be motivated to perform utility computation, ranking, and training repeatedly for each time period defined to refine the model training process, see e.g., KT[0065], where KT describes training a model against a training data set for several epochs to refine its results.
Regarding claim 18, this claim is similar in scope to claim 6.
Claim Rejections - 35 USC § 103
Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over Wang (US 11521090 B2) in view of You (Ranking and Tuning Pre-trained Models: A New Paradigm for Exploiting Model Hubs), and further in view of Moradi (US 20250013921 A1).
Regarding claim 10, Wang-You discloses “wherein the performing or the training of each of the corresponding FL ML models further comprises: select each of a number N of the FL ML models in a descending order of utility” (See [Page 14, Col 7, Lines 58 - 65]; Wang discloses performing a calculation of the utility of all edge nodes)
Wang-You fails to explicitly disclose, “select each of a number N of the FL ML models in a descending order of utility whilst maximizing: Utotal = U1 + U2 + … + UN where, Ui is the computed utility for the i-th FL ML model for 1 ≤ i ≤ N”.
Moradi teaches “select each of a number N of the FL ML models in a descending order of utility whilst maximizing: Utotal = U1 + U2 + … + UN where, Ui is the computed utility for the i-th FL ML model for 1 ≤ i ≤ N” (See [0091]; Moradi discloses obtaining the weights of all of the ML models taking a sum of the weights of all the ML models).
Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention having Wang-You and Moradi before them to modify Wang-You to compute a sum of all model utilities to obtain the total utility of all models. One would be motivated to calculate the total utility using a summation of model weights, which could include the utility of a model, see e.g., [0091], where Moradi calculates the model quality score using a sum of model weights and fine-tuned weights.
Claim Rejections - 35 USC § 103
Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over Wang (US 11521090 B2) in view of You (Ranking and Tuning Pre-trained Models: A New Paradigm for Exploiting Model Hubs), and further in view of Cornacchia (US 20230351477 A1).
Regarding claim 12,
Regarding claim 12, Wang discloses “train of the one or more FL ML models based on… a set of training samples associated with the FL ML model from local data of the edge device;” (See [Page 14, Col 7, Lines 10 - 14]; Wang discloses that there are training data samples at each node)
Wang-You fails to explicitly disclose, “train of the one or more FL ML models based on: obtain a set of training samples”
“apply a training sample to the FL ML model”
“estimate a loss based on a difference between the output of the FL ML model and the output indicated by each training sample”
“update a set of weights of the FL ML model based on the estimated loss”.
Cornacchia teaches “train of the one or more FL ML models based on: obtain a set of training samples” (See [0021]; Cornacchia discloses obtaining data from data gathered from merchants).
“apply a training sample to the FL ML model” (See [0062]; Cornacchia discloses training a model using a set of training data)
“estimate a loss based on a difference between the output of the FL ML model and the output indicated by each training sample” (See [0068]; Cornacchia discloses estimating a loss that determines a difference between output data generated by the model and the labels associated with the training data)
“update a set of weights of the FL ML model based on the estimated loss” (See [0100]; Cornacchia discloses updating a set of weights of the ML model)
Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention having Wang-You and Cornacchia before them to modify Wang-You to obtain a set of training samples, apply the samples to a model, estimate a loss between the model output and sample output, and update a set of weights in the model based on the estimated loss. One would be motivated to obtain a set of training samples and apply the training sample to the model to train the model using the obtained set of training samples. One would be motivated to estimate a loss between model output and training sample output to optimize a cost function, see e.g., [0068], where Cornacchia determines a loss between model output and training sample output to optimize a cost function. One would be motivated to update a set of model weights using the estimated loss to train the model multiple times, see e.g., [0100], where Cornacchia discloses updating a model’s weights using the loss by modifying one or more weights associated with the model numerous times as part of training the model.
Allowable Subject Matter
Claims 4 and 17 objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims. None of the prior art of recording, including Wang, You, Das, KT, Moradi, and Cornacchia, disclose the limitations of claims 4 and 17. Further, a complete and thorough search did not uncover any other prior art.
Claim 5 objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims. None of the prior art of recording, including Wang, You, Das, KT, Moradi, and Cornacchia, disclose the limitation of claim 5. Further, a complete and thorough search did not uncover any other prior art.
Claims 7 and 19 objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims. None of the prior art of recording, including Wang, You, Das, KT, Moradi, and Cornacchia, disclose the limitations of claims 7 and 19. Further, a complete and thorough search did not uncover any other prior art.
Claim 8 objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims. None of the prior art of recording, including Wang, You, Das, KT, Moradi, and Cornacchia, disclose the limitation of claim 8. Further, a complete and thorough search did not uncover any other prior art.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DAVID KIM whose telephone number is (571)272-4331. The examiner can normally be reached 7:30 AM - 4:30 PM.
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, Matthew Ell can be reached at (571) 270-3264. 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.
/D.K./Examiner, Art Unit 2141
/MATTHEW ELL/Supervisory Patent Examiner, Art Unit 2141
Prosecution Insights