DETAILED ACTION
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
This action is responsive to communications: RCE filed on 5/28/2026.
Claims 1, 7-8, 13-15, 17, 19-22, and 25-29 are pending. Claims 1, 15, and 29 are independent. Claims 2, 16, and 18 are newly canceled
The previous rejection of claim 1, 7-8, 13-15, 17, 19-22, and 25-29 under 35 USC § 103 have been withdrawn in view of the amendment.
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.
Claim(s) 1, 2, 7-8, 13, 15-22, 25, 27 and 29 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kruithof (“Object recognition using deep convolutional neural networks with complete transfer and partial frozen layers”) as made of reference in IDS dated 5/19/2021 in view of Whatmough et al. (US2020/0042877) and Jawahar et al. (US 10,776,693) in view of Gupta, (“Transfer learning & The art of using Pre-trained models in Deep Learning”)
In regards to claim 1, Kruithof substantially discloses a method for handling prediction of service characteristics using machine learning applied in a target domain, the method comprising:
Obtaining a source model for use in a source domain, wherein the source model was trained using observations collected in the source domain (Kruithof pg2 section2 para1, obtains a source model MS (base dataset A));
selecting a transfer configuration divides the source model into a first part and a second part (Kruithof pg2 section 1 para2, copy all layers and vary the number of layers that is fine-tuned); and
after obtaining the source model that was trained using observations collected in the source domain, creating a target model for use in the target domain(Kruithof Fig. 1 pg2 section2 para1, copying all layers…then freezing first N layers (fig. 1), network was pre-trained on A (source), then transferred and fine-tuned on B (target));
Kruithof does not explicitly disclose wherein creating the target model for use in the target domain comprises:
Collecting in the target domain a first set of observations; and
As a result of determining that the performance the candidate the candidate model does not the performance condition, performing further steps of:
Collecting in the target domain a second set of observations; and
Training the first modified second part of the source model using the second set of observations collected in the target domain, thereby producing a second modified second part of the source model; and
creating a second candidate model that comprises wherein a second candidate model comprises the first part of the source model and the second modified second part of the source model;
However Whatmough et al. discloses wherein creating the target model comprises:
Collecting in the target domain a first set of observations (Whatmough et al. para[0061], access a second set of data for retraining the model); and
As a result of determining that the performance the candidate the candidate model does not meet the performance condition (Whatmough et al. fig.534 para[0067], determination is made whether there are additional networks to be trained, if so a subsequent dataset is accessed), performing further steps of:
Collecting in the target domain a second set of observations (Whatmough et al. para[0067], subsequent data set is accessed); and
Training the first modified second part of the source model using the second set of observations collected in the target domain, thereby producing a second modified second part of the source model, wherein a second candidate model comprises the first part of the source model and the second modified second part of the source model (Whatmough et al. fig. 5 524 para[0065], trains second part (programmable layers) to create a subsequent neural net); and
Creating a second candidate model that comprises the first part of the source model and the second modified second part of the source model (Whatmough et al. fig. 6 618 para[0073], second maps such as concatenate F maps are provided to subsequent neural network layer).
It would have been obvious to one of ordinary skill in the art before the filing date of the invention to have combined the transfer learning method of Kruithof with the fixed hardware method of Whatmough et al. in order to provide power and performance advantages (Whatmough et al. para[0037]).
Kruithof does not explicitly disclose wherein creating the target model for use in the target domain comprises:
Training the second part of the source model, which was trained using observations collected in a source domain, using the first set of observations collected in the target domain, thereby producing a first modified second part of the source model;
Creating a first candidate model that comprises the first part of the source model and the first modified second part of the source model;
Determining whether a performance of the first candidate model meets a performance condition.
However Jawahar et al. substantially discloses wherein creating the target model for use in the target domain comprises:
Training the second part of the source model, which was trained using observations collected in a source domain, using the first set of observations collected in the target domain, thereby producing a first modified second part of the source model (Jawahar et al. fig. 4 406b, col22 ln20-49, The plurality of source specific features may be specific to the source domain and the plurality of common features may be common between the plurality of labeled text segments of the source domain and the plurality of unlabeled text segments of the target domain);
Creating a first candidate model that comprises the first part of the source model and the first modified second part of the source model (Jawahar et al. col23 fig. 4 col23 ln45 to col24 ln17, creates a first iteration using the common representation 406B and the target specific representation 416A);
Determining whether a performance of a first candidate model meets a performance condition (Jawahar et al. col20 ln37-50, may continue the iterative process of determining the target specific representation till the classification performance of the re-trained generalized classifier converges).
It would have been obvious to one of ordinary skill in the art before the filing date of the invention to have combined the transfer learning method of Kruithof with the domain adaptation method of Jawahar et al. in order to minimize the effects of domain specific features (Jawahar et al. col24 ln65 to col25 ln21).
Kruithof does not explicitly disclose determining a number of available observations in the target domain;
Based on the determined number of available observations in the target domain, selecting a set of candidate transfer configuration from a plurality of predefined sets of candidate transfer configuration;
selecting a transfer configuration from the selected set of candidate transfer configurations, wherein the selected transfer configuration.
However Gupta et al. discloses determining a number of available observations in the target domain (Gupta et al. pg12 scenario 2 and scenario 3, determine if the size of the data set is small or large);
Based on the determined number of available observations in the target domain, selecting a set of candidate transfer configuration from a plurality of predefined sets of candidate transfer configuration (Gupta et al. pg12 scenario 2 and scenario 3, based on size of dataset determine k number of layers to freeze, or if size of data is high enough retrain model from scratch);
selecting a transfer configuration from the selected set of candidate transfer configurations, wherein the selected transfer configuration (Gupta et al. pg12 scenario 2 and scenario 3, select transfer configuration from available scenarios).
It would have been obvious to one of ordinary skill in the art before the filing date of the invention to have combined transfer learning method of Kruithof with the pre-trained models of Gupta et al. in order to fine tune models to work in a target domain (Gupta et al. pg11 section “How can I use pre-trained models?” para4).
In regards to claim 7, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the method of claim 1, wherein
The source domain is a first data center (Jawahar et al. fig. 1 106 col7 ln33-62, The data processing server 104 may be configured to retrieve the labeled instances of the source domain from one or more social media websites or the database server106), and
The target domain is a second data center different than the first data center (Jawahar et al. fig. 1 102 col7 ln33-62, the data processing server 104 may be configured to receive the classification request from the user-computing device 102 for classification of the plurality of unlabeled instances of the target domain).
It would have been obvious to one of ordinary skill in the art before the filing date of the invention to have combined the transfer learning method of Kruithof with the domain adaptation method of Jawahar et al. in order to minimize the effects of domain specific features (Jawahar et al. col24 ln65 to col25 ln21).
In regards to claim 8, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the method of claim 1, wherein the observations include measurements and samples taken in the source and target domains, respectively (Kruithof pg3 section3.2 para1).
In regards to claim 13, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the method of claim 1, wherein the observations are related to performance of the service, and/or to current usage of processing and storing resources (Kruithof pg1 abstract, allow flexible querying in a large number of cameras, especially for security applications)
.
Claim 15 recites substantially similar limitations to claim 1. Thus claim 15 is rejected along the same rationale as claim 1.
In regards to claim 17, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the machine learning manager of claim 15, wherein the machine learning manager is configured to select the transfer configuration by training the second part according to the selected set of candidate transfer configurations and by selecting the candidate transfer configuration from the selected set of candidate transfer configurations that provides the most accurate target model (Kruithof fig.2 pg6 section4 para1).
In regards to claim 19, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the machine learning manager of claim 17, wherein the machine learning manager is configured to select the transfer configuration by evaluating the candidate transfer configurations with respect to one or more predefined criteria (Kruithof pg2section1 para2).
In regards to claim 20, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the machine learning manager of claim 19, wherein the one or more predefined criteria is/are configured to select the candidate transfer configuration that provides a target model with the highest accuracy and/or lowest error (Kruithof pg2section1 para2).
In regards to claim 21, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the machine learning manager of claim 15, wherein the source and target domains refer to different sets of computing resources and/or different prediction tasks (Kruithof pg6 section3.4 para5).
In regards to claim 22, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the machine learning manager of claim 15, wherein the observations include measurements and samples taken in the source and target domains, respectively (Kruithof pg3 section3.2 para1).
In regards to claim 25, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the machine learning manager of claim 15, wherein the source model and the target model are based on a neural network where the first part of the source model comprises a set of initial weights in the neural network and the second part of the source model comprises a set of subsequent weights in the neural network (Kruithof fig. 1 pg3 section3.1 para3).
In regards to claim 27, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the machine learning manager of claim 15, wherein said observations are related to performance of the service and/or to current usage of processing and storing resources (Kruithof pg1 abstract, allow flexible querying in a large number of cameras, especially for security applications).
In regards to claim 29, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. discloses a computer program product comprising a non-transitory computer readable medium storing a computer program comprising instructions which, when executed on at least one processor, cause the at least one processor to carry out the method of claim1 (Kruithof pg2 section1 para2).
Claim(s) 14 and 28 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kruithof in view of Whatmough et al., Jawahar et al., Gupta et al. and Dias (US2019/0303211).
In regards to claim 14, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. discloses the method of claim 1.
Kruithof does not explicitly disclose wherein the method comprises using the target model to predict whether a Service Level Agreement has been violated in the target domain.
However Dias discloses wherein the method comprises using the target model to predict whether a Service Level Agreement has been violated in the target domain (Dias para[0032]).
It would have been obvious to one of ordinary skill in the art before the filing date of the invention to have combined the transfer learning method with the resource allocation method of Dias in order to maximize expected profits and avoid violating agreements (Dias para[0023]).
In regards to claim 28, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. discloses the machine learning manager of claim 15.
Kruithof does not explicitly disclose wherein the prediction of service characteristics in the target domain comprises predicting whether a Service Level Agreement, has been violated in the target domain.
However Dias discloses wherein the prediction of service characteristics in the target domain comprises predicting whether a Service Level Agreement, has been violated in the target domain (Dias para[0032]).
It would have been obvious to one of ordinary skill in the art before the filing date of the invention to have combined the transfer learning method with the resource allocation method of Dias in order to maximize expected profits and avoid violating agreements (Dias para[0023]).
Claim(s) 26 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kruithof in view of Whatmough et al., Jawahar et al., Gupta et al. and Wang (US2019/0325621).
In regards to claim 26, Kruithof as modified by Whatmough et al., Jawahar et al. and Gupta et al. disclose the machine learning manager of claim 15.
Kruithof does not explicitly disclose wherein the source model and target model comprise a random-forest model with a number of trees where the first part of the source model comprises a first set of trees and the second part of the source model comprises a second set of trees.
However Wang substantially discloses disclose wherein the source model and target model comprise a random-forest model with a number of trees where the first part of the source model comprises a first set of trees and the second part of the source model comprises a second set of trees (Wang para[0119]).
It would have been obvious to one of ordinary skill in the art before the filing date of the invention to have combined the transfer learning method of Kruithof with the Random Forrest classifier method of Wang in order to improve accuracy of models (Wang et al. para[0124]).
Response to Arguments
Applicant's arguments filed 5 have been fully considered but they are not persuasive. Applicant argues on pg7 that Kruithof does not teach “based on a determined number of available observations in the target domain, selecting a set of candidate transfer configurations from a plurality of predefined sets of candidate transfer configurations”
However Kruithof as modified by Whatmough et al. and Jawahar et al. and Gupta et al. discloses based on a determined number of available observations in the target domain, selecting a set of candidate transfer configurations from a plurality of predefined sets of candidate transfer configurations (Gupta et al. pg12 scenario 2 and scenario 3, in a scenario with a small set of data you would freeze k initial layers and retrain the others, in a scenario with a large data set train the model from scratch, Kruithof fig. 4 pg4 section 3.3 shows accuracy as function of the sized of the target data set and the number of fixed layers. It would have been obvious to one of ordinary skill in the art to select a transfer configuration that would provide the highest accuracy based on the available target dataset).
Applicant argues on pg8 that Kruithof does not teach “as a result of determining that the performance of the first candidate model does not meet the performance condition performing the further steps of collecting in the target domain a second set of observations; training the first modified second part of the source model using the second set of observations collected in the target domain, thereby producing a second modified second part of the source model; and creating a second candidate model comprising the first part of the source model and the second modified second part of the source model”.
However Kruithof as modified by Whatmough et al., Jawahar et al., and Gupta et al. discloses as a result of determining that the performance of the first candidate model does not meet the performance condition performing the further steps of collecting in the target domain a second set of observations (Whatmough et al. para[0061], collects data from a subsequent data set to be used to generate a subsequent set of weights); training the first modified second part of the source model using the second set of observations collected in the target domain, thereby producing a second modified second part of the source model (Whatmough et al. fig. 5 524 para[0065], the second programable portion is retrained by the subsequent data); and creating a second candidate model comprising the first part of the source model and the second modified second part of the source mode (Whatmough et al. fig. 6 618 para[0073], concatenate first and second portion to generate neural network).
Applicant argues on page11 that Kruithof does not teach “creating the target model for use in the target domain comprises:…training the second part of the source model, which was trained using observations collected in a source domain, thereby producing a first modified second part of the source model”
However Kruithof as modified by Whatmough et al., Jawahar et al., and Gupta et al. discloses creating the target model for use in the target domain comprises:…training the second part of the source model, which was trained using observations collected in a target domain, thereby producing a first modified second part of the source model (Jawahar et al. fig. 4 406b, col22 ln20-49, after receipt of the real-time input data, the data processing server may be configured to formulate the first neural network for the classification of the plurality of unlabeled text segments associated with the target domain).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to NICHOLAS HASTY whose telephone number is (571)270-7775. The examiner can normally be reached Monday-Friday 8:30am-5:00pm.
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, Matt 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.
/N.H/Examiner, Art Unit 2141
/MATTHEW ELL/Supervisory Patent Examiner, Art Unit 2141
Prosecution Insights