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 .
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and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicants’ submission filed on 9/29/25 has been entered.
DETAILED ACTION
The instant application having Application No. 17455241 has a total of 20 claims pending in the application.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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 1-2, 6-9, 13-16, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Dong et al (US 20160358095 A1) in view of Chen (US 20180314971 A1), Huo et al (CN 105378762 B, see English translation attached), and Zhang et al (“Adaptive Client Selection in Resource Constrained Federated Learning Systems: A Deep Reinforcement Learning Approach”).
As per claims 1, 8, and 15, Dong et al discloses, “A computer implemented method” (pg.3, particularly paragraph 0034-0037; EN: this denotes the hardware for the system). “for training of machine learning (ML) models” (Pg.4, particularly paragraph 0045; EN: this denotes training a decision tree). “and executing inference using the ML models” (Pg.4, particularly paragraph 0044; EN: this denotes using the tree for predictions). “… the method being executed by one or more processors and comprising” (pg.3, particularly paragraph 0034-0037; EN: this denotes the hardware for the system).
“transmitting, from a resource manager node, a set of training tasks to a set of worker nodes” (pg.3-4, particularly paragraph 0043; EN: this denotes a Source Processing item (PI) distributing data to a plurality of local statistic PIs that perform local statistic for decision tree learning (i.e. training tasks). Here the local PIs are the worker nodes, and the resource manager node is the source PI). “The set of workers nodes comprising two or more worker nodes distributed across the … system and each receiving data from a respective local data store, local data stores being distributed across the … system with respective worker nodes” (pg.3-4, particularly paragraph 0043; EN: this denotes a Source Processing item (PI) distributing data to a plurality of local statistic PIs that perform local statistic for decision tree learning (i.e. training tasks. Each processor will inherently have some sort of data store to store data being processed by the processor). “Each local data store storing a respective sub-set of features received from , the subsets of features between local data stores each comprising two or more features” (pg.4, particularly paragraph 0049; EN: this denotes distributing the plurality of data instances to the PI’s).
“the set of training tasks being executed to provide a ML model” (Pg.3, particularly paraph 0042; EN: this denotes using this to create a global decision tree).
“by each worker node in the set of worker nodes:” ” (pg.3-4, particularly paragraph 0043; EN: this denotes a Source Processing item (PI) distributing data to a plurality of local statistic PIs).
“executing a respective training task to provide a set of local parameters” (pg.3-4, particularly paragraph 0043; EN: this denotes a Source Processing item (PI) distributing data to a plurality of local statistic PIs that perform local statistic for decision tree learning (i.e. training tasks).
“transmitting the set of local parameters to the … node” (pg.3-4, particularly paragraph 0043; EN: this denotes the local statistic PI passing it to a model aggregator PI to aggregate the data from the local statistic PIs).
“receiving sets of local parameters, each set of local parameters being received from a respective worker node…” (pg.3-4, particularly paragraph 0043; EN: this denotes the local statistic PI passing it to a model aggregator PI to aggregate the data from the local statistic PIs).
“Merging, by the … node, two or more sets of local parameters to provide a set of global parameters” (pg.3, particularly paragraph 0042; EN: this denotes the aggregating of local statistics to a global decision tree).
“receiving …a set of local optimal splits each local optimal split in the set of local optimal splits…” (Pg.4, particularly paragraph 0049; EN: this denotes using local model PIs to do local processing of splits for the tree).
“determining an optimal global split based on the set of local optimal splits” (Pg.4, particularly paragraph 0050; EN: this denotes sending the local splits to conflict resolve PIs and pick the best split from the local PIs). “the optimal global split representing a feature of the ML model” (pg.4, particularly paragraph 0045; EN: this denotes the use of information gain to select splits, with the attribute being used the feature of the algorithm).
“updating by the … node, the ML model based on the optimal global split” (Pg.4, particularly paragraph 0050; EN: this denotes using local model PIs to do local processing of splits for the tree and update the tree from the conflict resolution PIs).
However, Dong fails to explicitly disclose, “in cloud systems”, “the sub-sets of features between local data stores each comprising two or more features and being non-overlapping” “transmitting the set of parameters to the resource manager node”, “… by the resource manager node…”, “…and being associated with a response time”, “selecting two or more sets of parameters using response times, the two or more sets of locale parameters being less than all sets of local parameters received from the set of worker nodes”, “transmitting, by the resource manager node, a subset of global parameters to each parameter server in a set of parameter servers”, “receiving, by the resource manager node… being transmitted to the resource manager node from a respective parameter server”, “… by the resource manager node….”
Chen discloses, “in cloud systems” (Pg.2, particularly paragraph 0025; EN: this denotes the system being implemented on a cloud).
“transmitting the set of parameters to the resource manager node”, “… by the resource manager node…” and “… by the resource manager node…” (Pg.2-3, particularly paragraph 0027; EN: this denotes the master processor sending the data to the worker groups and receiving the final results back. The Dong reference has different processors for sending initial data and receiving final results. The Chen reference shows that it is known in the art to use a single processor both to send initial data and receive final results).
“transmitting, by the resource manager node, a subset of global parameters to each parameter server in a set of parameter servers” and “receiving, by the resource manager node… being transmitted to the resource manager node from a respective parameter server” (Pg.2-3, particularly paragraph 0027; EN: this denotes having training groups which can be assigned as needed to different machine learning processes. When combined with the Chen reference, this shows that different processing groups can perform different machine learning processes, such as the various split determination and training aspects of the Dong reference).
Huo discloses, “the sub-sets of features between local data stores each comprising two or more features and being non-overlapping” (Pg.5, fifth paragraph EN: This denotes breaking up the training data into subsets that are non-overlapping for distributed processing).
Zhang discloses, “…and being associated with a response time”, “selecting two or more sets of parameters using response times, the two or more sets of locale parameters being less than all sets of local parameters received from the set of worker nodes” (Pg.98425-98426, particularly Section B; EN: this denotes selecting clients to update the model based on wireless bandwidth and the time required to update the model using those clients).
Dong and Chen are analogous art because both involve distributed machine learning.
Before the effective filing date it would have been obvious to one skilled in the art of distributed machine learning to combine the work of Dong and Chen in order to use different distributions of processors for machine learning tasks.
The motivation for doing so would be to allow “training jobs may be ordered at different times. Accordingly, the allocation ... of the training jobs may occur over time” (Chen, Pg.3, paragraph 0028) or in the case of Dong, allow the system to assign out tasks to worker processors as needed over time by the system in the manner most efficient for the system.
Therefore before the effective filing date it would have been obvious to one skilled in the art of distributed machine learning to combine the work of Dong and Chen in order to use different distributions of processors for machine learning tasks.
Dong and Huo are analogous art because both involve distributed machine learning.
Before the effective filing date it would have been obvious to one skilled in the art of distributed machine learning to combine the work of Dong and Chen in order to have non-overlapping training data distributed.
The motivation for doing so would be to make sure that “each data block keeps the joint distribution of input character and output types, namely p (input, output)” (Huo, Pg.5, third paragraph) or in the case of Dong, allow the system to distribute the training data equally out to the distributed processors as needed.
Therefore before the effective filing date it would have been obvious to one skilled in the art of distributed machine learning to combine the work of Dong and Chen in order to have non-overlapping training data distributed.
Dong and Zhang are analogous art because both involve distributed machine learning.
Before the effective filing date it would have been obvious to one skilled in the art of distributed machine learning to combine the work of Dong and Zhang in order to consider time delay when selecting clients to update the global model.
The motivation for doing so would be to allow the system to select clients effectively because “Due to the limited wireless bandwidth and energy of mobile devices, it is not practical for FL to perform model updating and aggregation on all participating devices in parallel. And it is difficulty for FL server to select apposite clients to take part in model training which is important to save energy and reduce latency” (Zhang, Abstract) or in the case of Dong, allow the system to select clients that can respond in adequate time to optimize updating the model.
Therefore before the effective filing date it would have been obvious to one skilled in the art of distributed machine learning to combine the work of Dong and Zhang in order to consider time delay when selecting clients to update the global model.
As per claim 2, 9, and 16, Dong discloses, “Wherein each training task comprises a set of identifiers, each identifier indicating data that is to be used to execute the training task” (Pg.4, particularly paragraph 0042; EN: this denotes designating data instances, which is the identifier, and passing those out to the workers to be executed).
As per claims 6, 13 and 20, Dong discloses, “wherein each local data store comprises a column-oriented data store” (Fig. 4 and associated paragraphs; NE: this denotes the data being represented as a set of rows and columns).
As per claims 7 and 14, Dong discloses, “wherein the ML model comprises a decision tree” (pg.3, particularly paragraph 0042; EN: this denotes the ML model being a decision tree).
Claim Rejections - 35 USC § 103
Claims 3, 10 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Dong et al (US 20160358095 A1) in view of Chen (US 20180314971 A1), Huo et al (CN 105378762 B, see English translation attached), and Zhang et al (“Adaptive Client Selection in Resource Constrained Federated Learning Systems: A Deep Reinforcement Learning Approach”) as shown for claims 1, 8 and 15 above and further in view of Ebrahimi et al (US 20200372337 A1).
As per claims 3, 10, and 17, Dong fails to explicitly disclose, “wherein the two or more sets of local parameters is determined based on a control parameter that limits the two or more sets of local parameters to less than all sets of local parameters received from the set of worker nodes”
Ebrahimi discloses, “wherein the two or more sets of local parameters is determined based on a control parameter that limits the two or more sets of local parameters to less than all sets of local parameters received from the set of worker nodes” (Pg.4, particularly paragraph 0036; EN: this denotes the use of batch size, which denotes how much data is used at once. The batch size here does not use all of the data at once, but limits it to a certain amount of data per input).
Dong and Ebrahimi are analogous art because both involve distributed machine learning.
Before the effective filing date it would have been obvious to one skilled in the art of distributed machine learning to combine the work of Dong and Ebrahimi in order to use only a certain amount of local data at once when processing at the global model.
The motivation for doing so would be because “When global batch size is scaled, [it] improves end to end training times associated with training a neural network” (Ebrahimi, Pg.4, paragraph 0036) or in the case of Dong, allow the system to have a batch size appropriate to its needs to make the training/aggregation process of the machine learning algorithm efficient.
Therefore before the effective filing date it would have been obvious to one skilled in the art of distributed machine learning to combine the work of Dong and Ebrahimi in order to use only a certain amount of local data at once when processing at the global model.
Claim Rejections - 35 USC § 103
Claims 4-5, 11-12, and 18-19 are rejected under 35 U.S.C. 103 as being unpatentable over Dong et al (US 20160358095 A1) in view of Chen (US 20180314971 A1), Huo et al (CN 105378762 B, see English translation attached), and Zhang et al (“Adaptive Client Selection in Resource Constrained Federated Learning Systems: A Deep Reinforcement Learning Approach”) as shown for claims 1, 8 and 15 above and further in view of El-Khamy et al (US 20210374608 A1).
As per claims 4, 11, and 18, Dong discloses, “further comprising, during inference” (Pg.4, particularly paragraph 0044; EN: this denotes using the tree for predictions).
“receiving, by a worker node, the ML model form the … node” (Pg.4, particularly paragraph 0052; EN: this denotes the Pi having access to a replica of the decision tree).
“determining, by the worker node, a … based on the ML model and data stored in a local data store accessible by the worker node” (Pg.4, particularly paragraph 0053; EN: this denotes the processing of the worker node on its data. The worker node inherently has some form of memory to interact with the data it is working on).
“Transmitting, by the worker node, the … to the … node” (Pg.4, particularly paragraph 0053; EN: this denotes transferring the determined data to the conflict PI to analyze the changes).
“Determining, …, an inference result of the ML model at least partially based on the result” (Pg.4, particularly paragraph 0044; EN: this denotes using the tree for predictions).
Chen discloses, “From the resource manager node”, “a binary code based on the ML model”, “transmitting, by the worker node, the binary code to the resource manager”, “Providing, by the resource manager node, a … to a parameter server, …, the parameter server executing an operation on … to provide a result to the to the source manager”, and “by the resource manager” (Pg.2-3, particularly paragraph 0027; EN: this denotes having training groups which can be assigned as needed to different machine learning processes. When combined with the Chen reference, this shows that different processing groups can perform different machine learning processes, such as the various split determination and training aspects of the Dong reference).
Dong fails to explicitly disclose, “A binary code based on the ML model”, “the binary code”, “a set of binary codes” “the set of binary codes comprising the binary code” “the set of binary codes.”
El-Khamy discloses, “A binary code based on the ML model”, “the binary code”, “a set of binary codes” “the set of binary codes comprising the binary code” “the set of binary codes” (Pg.6, particularly paragraph 0040; EN: this denotes using a binary vector to represent local parameters).
Dong and El-Khamy are analogous art because both involve distributed machine learning.
Before the effective filing date it would have been obvious to one skilled in the art of distributed machine learning to combine the work of Dong and El-Khamy in order to use binary to represent data.
The motivation for doing so would be to use “a binary vector that indicates active factors” (Pg.4, particularly paragraph 0032) or in the case of Dong, allow the system to use binary to represent what data is being passed around and used by the system.
Therefore before the effective filing date it would have been obvious to one skilled in the art of distributed machine learning to combine the work of Dong and El-Khamy in order to use binary to represent data.
As per claims 5, 12, and 19, El-Khamy discloses, “Wherein the binary code represents a portion of the ML model that the worker node is capable of resolving using at least a portion of the data stored in the local data store accessible by the worker” (Pg.6, particularly paragraph 0040; EN: this denotes using a binary vector to represent local parameters).
Response to Arguments
Applicant's arguments with respect to claims 1-20 have been considered but are moot in view of the new ground(s) of rejection.
Conclusion
The examiner requests, in response to this Office action, support be shown for language added to any original claims on amendment and any new claims. That is, indicate support for newly added claim language by specifically pointing to page(s) and line no(s) in the specification and/or drawing figure(s). This will assist the examiner in prosecuting the application.
When responding to this office action, Applicant is advised to clearly point out the patentable novelty which he or she thinks the claims present, in view of the state of the art disclosed by the references cited or the objections made. He or she must also show how the amendments avoid such references or objections See 37 CFR 1.111(c).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to BEN M RIFKIN whose telephone number is (571)272-9768. The examiner can normally be reached Monday-Friday 9 am - 5 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, Alexey Shmatov can be reached at (571) 270-3428. 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.
/BEN M RIFKIN/Primary Examiner, Art Unit 2123