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 .
Response to Arguments
Applicant's arguments filed 6/17/2026 have been fully considered but they are not persuasive.
Applicant argues, “Lu fails to disclose determining the pre-training task which is currently traversed as the current pre-training task in each round of training, when traversing cyclically the M pre-training tasks until the model converges.” Remarks 10. Applicant claims, “traversing cyclically the M pre-training tasks until the model converges, and in each round of training, determining the pre-training task which is currently traversed as the current pre-training task…” Claim 1. Traversing cyclically is not a term of art and it is not defined in the specification. Lu sec. 3.3 teaches the claimed element as “a round-robin batch-level sampling regime that cycles through each task from the beginning of multi-task training. As such, one multi-task iteration consists of each task forwarding a batch and updating parameters in sequence.”
Applicant argues, “Lu also fails to disclose acquiring a loss function corresponding to the determined current pre-training task…” Remarks 10. Lu sec. 3.3 teaches a loss function because that’s what training is – updating weights based on a loss function – and Lu teaches that the training is done on each task, “We consider a round-robin batch-level sampling regime that cycles through each task from the beginning of multi-task training. As such, one multi-task iteration consists of each task forwarding a batch and updating parameters in sequence.”
Applicant argues, “Lu also fails to disclose … updating model parameters corresponding to the determined current pre-training task according to the loss function.
Remarks 10.
(Lu sec. 3.3 teaches exactly this idea when it teaches training and when it says, about its training, “As such, one multi-task iteration consists of each task forwarding a batch and updating parameters in sequence.” The parameters are model parameters being updated based on a loss function for the task that is being trained.
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
Claims 1, 4-9, 12-16 and 19-20 are rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claims contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventors, at the time the application was filed, had possession of the claimed invention. Applicant claims, “traversing cyclically the M pre-training tasks until the model converges, and in each round of training, determining the pre-training task which is currently traversed as the current pre-training task…” This is not described in the specification. As a whole, it is not described, but also traversing cyclically is not described.
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, 4-9, 12-16 and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over The Natural Language Decathlon: Multitask Learning as Question Answering by McCann et al, US20190384304A1 by Towal et al, Multi Output Learning using Task Wise Attention for Predicting Binary Properties of Tweets: Shared-Task-On-Fighting the COVID-19 Infodemic by Suhane et al and 12-in-1: Multi-Task Vision and Language Representation Learning by Lu et al.
McCann teaches claims 1, 9 and 16. (Currently Amended) A computer-implemented method for acquiring a pre-trained model, comprising: (McCann abs “new multitask question answering network (MQAN) that jointly learns all tasks in decaNLP without any task-specific modules or parameters. MQAN shows improvements in transfer learning…” McCann p. 20 equations 18 is computed. That means the MQAN is a computer implemented method.)
acquiring a pre-training task set composed of M pre-training tasks, M being a positive integer greater than 1, the pre-training tasks comprising: N question-answering tasks corresponding to different question-answering forms, N being a positive integer greater than 1 and less than or equal to M, the different question-answering forms comprise text question-answering, (McCann Fig 1. Questions and answers in general.) knowledge-based question-answering, (McCann Fig. 1 questions and answers in general) table question-answering, (McCann fig. 1 shows a question with a SQL table as context.) image question-answering (McCann p. 9 “These tasks can also be learned with image classification… MQAN trained on decaNLP is the first, single model to achieve reasonable performance on such a wide variety of complex NLP tasks without task-specific modules or parameters…”) and video question-answering; (McCann fig. 1 shows that videos/films such as Harry potter and Beuty and the Beast can serve as context for the question answering tasks.) and (McCann p.2 sec. 2 “decaNLP consists of 10 publicly available datasets with examples cast as (question, context, answer) triplets as shown in Fig. 1.”)
acquiring the pre-trained model by jointly training according to the M pre-training tasks, (McCann p. 4 sec. 3 “Because every task is framed as question answering and trained jointly, we call our model a multitask question answering network (MQAN). Each example consists of a context, question, and answer as shown in Fig. 1.”)
wherein the step of acquiring the pre-trained model by jointly training according to the M pre-training tasks comprises:
performing the following processing respectively in each round of training:
determining the pre-training task corresponding to the round of training as a current pre-training task,(McCann p. 7 sec. 4.2 “curriculum learning jointly trains the easier tasks (SST, QA-SRL, QA-ZRE, WOZ, WikiSQL, and MWSC) first.”)
acquiring a loss function corresponding to the determined current pre-training task; and (McCann p. 6 sec. 3 “We train using a token-level negative log-likelihood loss over all time-steps…”)
updating model parameters corresponding to the determined current pre-training task (McCann p. 5 between eqn 7 and 8 “This intermediate state is used to get attention weights α C t and α Q t to allow the decoder to focus on encoded information relevant to time step t.”)
wherein each of the M pre-training tasks is taken as the current pre-training task, (McCann p. 7 sec. 4.2 “curriculum learning jointly trains the easier tasks (SST, QA-SRL, QA-ZRE, WOZ, WikiSQL, and MWSC) first.”)
.
McCann doesn’t say that parameters are updated base on loss functions.
However, Towal teaches updating model parameters corresponding to the current pre-training task according to the loss function. (Towal para 153 “the loss functions may be combined to form a total loss, and the total loss may be used to train (e.g., update the parameters of) the machine learning model(s) 104.”)
the step of acquiring a loss function corresponding to the current pre-training task comprises: acquiring L loss functions corresponding to the current pre-training task, L being a positive integer; and (Towal para 153 “the path geometry(ies) 106 predicted by the machine learning model(s) 104 may use a first loss function and the path type(s) 108 may use a second loss function (e.g., different from the first loss function).”)
when L is greater than 1, the step of updating model parameters corresponding to the current pre-training task according to the loss function comprises: determining a comprehensive loss function according (Towal para 153 “the loss functions may be combined to form a total loss,” Towal uses more than one loss function.) and updating the model parameters corresponding to the current pre-training task according to the comprehensive loss function, (Towal para 153 “and the total loss may be used to train (e.g., update the parameters of) the machine learning model(s) 104.”)
Towal, McCann and the claims all train neural networks. It would have been obvious to a person having ordinary skill in the art, at the time of filing, to update parameters based on multiple loss functions “where conventional approaches are unreliable or would otherwise fail…” Towal para 8.
Towal and McCann don’t teach a weighted summation of the loss functions.
However, Suhane teaches a comprehensive loss function according to a weight summation result of the L loss functions. (Suhane p. 113 left col. Mid-page “The loss function was weighted-cross-entropy for each task, and the final loss was the sum of losses for the 7 tasks.”)
Towal, McCann, Suhane and the claims are all directed to multi-loss calculations for tasks. It would have been obvious to a person having ordinary skill in the art, at the time of filing, to weight the loss functions in the summed comprehensive loss function because it is a “natural approach” when “input data is skewed in the distribution of labels for each question.” Suhane p. 112 sec. 3.4. This skewed label distribution is described in fig. 3 of McCann, “When p(generation) is highest, the MQAN places the most weight on the external vocab. When p(context) is highest, the MQAN places the most weight on the pointer distribution over the context. When p(question) is highest, the MQAN places the most weight on the pointer distribution over the question.”
Towal doesn’t teach cyclical pretraining.
However, Lu teaches the M pre-training tasks are traversed cyclically until the model converges, and in each round of training, the pre-training task which is currently traversed is taken as the current pre-training task. (Lu sec. 3.3 “We consider a round-robin batch-level sampling regime that cycles through each task from the beginning of multi-task training. As such, one multi-task iteration consists of each task forwarding a batch and updating parameters in sequence.” Lu Sec. 3.3 “If performance improvement is less than 0.1% over 2 epochs, we consider it Converged and shift it into stop mode.”)
Towal, Lu and the claims are all directed to multiple loss calculations. It would have been obvious to a person having ordinary skill in the art, at the time of filing, to cyclically traverse multiple pre-training tasks in order to add Lu’s round-robin training to overcome excessive training iterations sometimes caused by multitask training (Lu sec. 3.3 “84k iterations” for multi task v. “5k iterations” for single task) and stopping training on convergence to avoid “drastically overtrain[ing] on smaller tasks leading to overfitting.” Lu sec. 3.3.
McCann teaches claims 4, 12 and 19. The method according to claim 1, wherein the pre-training task set comprises: a question-answering pre-training task subset; and (McCann p.2 sec. 2 “decaNLP consists of 10 publicly available datasets with examples cast as (question, context, answer) triplets as shown in Fig. 1.”)
the question-answering pre-training task subset comprises: the N question-answering tasks, (McCann p.2 sec. 2 “decaNLP consists of 10 publicly available datasets with examples cast as (question, context, answer) triplets as shown in Fig. 1.”) and one or any combination of the following: a task of judging matching between a question and a data source, a task of detecting a part related to the question in the data source, (McCann p. 5 between eqn. 8 and 9 “Context representations are combined with these weights and fed through a feedforward network with tanh activation to form the recurrent context state…” The context state is the detected part related to the question in the data source.) and a task of judging validity of the question and/or the data source.
McCann teaches claims 5, 13 and 20. The method according to claim 1, wherein the pre-training task set further comprises one or all of the following: a single-mode pre-training task subset and a multi-mode pre-training task subset; and
the single-mode pre-training task subset comprises: P different single-mode pre-training tasks, P being a positive integer; and the multi-mode pre-training task subset comprises: Q different multi-mode pre-training tasks, Q being a positive integer. (McCann p. 6 table 2 “Multitask models use a round-robin batch-level sampling strategy to jointly train on the full decaNLP. The last column includes an additional anti-curriculum (+ACurr) phase that trains on SQuAD alone before switching to the fully joint strategy.” Training on SQuAD alone is the P single-mode training, and jointly training is the multi-mode training.)
McCann teaches claims 6 and 14. The method according to claim 2, wherein
the pre-training task set further comprises one or all of the following: a single-mode pre-training task subset and a multi-mode pre-training task subset; and
the single-mode pre-training task subset comprises: P different single-mode pre-training tasks, P being a positive integer; and the multi-mode pre-training task subset comprises: Q different multi-mode pre-training tasks, Q being a positive integer. (McCann p. 6 table 2 “Multitask models use a round-robin batch-level sampling strategy to jointly train on the full decaNLP. The last column includes an additional anti-curriculum (+ACurr) phase that trains on SQuAD alone before switching to the fully joint strategy.” Training on SQuAD alone is the P single-mode training, and jointly training is the multi-mode training.)
McCann teaches claims 7 and 15. The method according to claim 1, wherein the pre-training task set further comprises one or all of the following: a single-mode pre-training task subset and a multi-mode pre-training task subset; and
the single-mode pre-training task subset comprises: P different single-mode pre-training tasks, P being a positive integer; and the multi-mode pre-training task subset comprises: Q different multi-mode pre-training tasks, Q being a positive integer. (McCann p. 6 table 2 “Multitask models use a round-robin batch-level sampling strategy to jointly train on the full decaNLP. The last column includes an additional anti-curriculum (+ACurr) phase that trains on SQuAD alone before switching to the fully joint strategy.” Training on SQuAD alone is the P single-mode training, and jointly training is the multi-mode training.)
McCann teaches claim 8. The method according to claim 4, wherein the pre-training task set further comprises one or all of the following: a single-mode pre-training task subset and a multi-mode pre-training task subset; and
the single-mode pre-training task subset comprises: P different single-mode pre-training tasks, P being a positive integer; and the multi-mode pre-training task subset comprises: Q different multi-mode pre-training tasks, Q being a positive integer. (McCann p. 6 table 2 “Multitask models use a round-robin batch-level sampling strategy to jointly train on the full decaNLP. The last column includes an additional anti-curriculum (+ACurr) phase that trains on SQuAD alone before switching to the fully joint strategy.” Training on SQuAD alone is the P single-mode training, and jointly training is the multi-mode training.)
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Austin Hicks whose telephone number is (571)270-3377. The examiner can normally be reached Monday - Thursday 8-4 PST.
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, Mariela Reyes can be reached at (571) 270-1006. 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.
/AUSTIN HICKS/Primary Examiner, Art Unit 2142
Prosecution Insights