DETAILED ACTION
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 § 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.
Claim(s) 1-2, 5, 8, 15, and 18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Bacco, Luca, et al. "Explainable sentiment analysis: a hierarchical transformer-based extractive summarization approach." in view of Kokalj, Enja, et al. "BERT meets shapley: Extending SHAP explanations to transformer-based classifiers." further in view of Cheng et al. US 2022/0405623.
Regarding claims 1, 2, and 12, Bacco teaches “a system for using explainability vectors to generate contextual data for training downstream models, the system comprising: receiving training data for an upstream machine learning model” (pg. 8 figure 1
PNG
media_image1.png
690
722
media_image1.png
Greyscale
shows transformer 1 and transformer 2, an upstream and downstream model), wherein the training data comprises values for a first set of features” (pg. 6 §3.1 “To benchmark our models, we used the IMDB Large Movie Review Dataset […] The data are already divided into two equivalent sets, one for training and one for testing”);
“training the upstream machine learning model based on the training data, wherein the upstream machine learning model is a bidirectional encoder representation transformer model which outputs text representations” (pg. 9 §3.2.3 “T1: For a fair comparison, the first transformer model was the same for both the architectures; we opted to use the pre-trained version of RoBERTa [36];” RoBERTa is a bidirectional encoder representation transformer model), “and wherein the text representations generated by the upstream machine learning model are input to a downstream machine learning model” (pg. 7 §3.2.1 “The input of the first transformer is a sequence of t tokens, while the output is an embedding representation of that sequence.” which as shown in figure 1 is input into transformer 2, “After T1 has elaborated the N sequences, the new generated representations {r1...rN} are stacked together to become the input of T2.”);
“using the upstream machine learning model, generating a first set of outputs based on the training data” (previous citation “The input of the first transformer is a sequence of t tokens, while the output is an embedding representation of that sequence.”);
The Bacco reference has been addressed above. More specifically, Kokalj teaches “processing the upstream machine learning model to extract an explainability vector, wherein each entry in the explainability vector corresponds to a feature in the first set of features and is indicative of a correlation between the feature and output of the upstream machine learning model” (Kokalj pg. 3 right col. “Finally, the predictions for the new locally generated instances are produced and returned to the Kernel SHAP explainer. With this modification, SHAP is able to compute the features’ impact on the prediction (i.e. the explanation)”);
“processing the second set of features and the first set of outputs to generate an explanative factor, wherein the explanative factor includes a set of features specifying real values which correspond to correlations between the second set of features and the first set of outputs” (Kokalj figure 4 shows the real values and correlations, pg. 4 last ¶ “In this way, the graph allows the user to compare the direction of the impact (positive/negative) and also the magnitude of impact for individual words. The bottom text box representation of the sentence shows the words colored green if they significantly contributed to the prediction and red if they significantly opposed it”);
It would have been obvious to one having ordinary skill in the art at the time that the invention was effectively filed to combine the teachings of Bacco with that of Kokalj since a combination of known methods would yield predictable results. As shown in Kokalj, it is known the generate explanation at the end of a model using various techniques. Therefore these techniques would operate in a known and predictable manner with the system above.
The Bacco and Kokalj references have been addressed above. They do not explicitly teach correlation thresholds. Cheng however teaches “based on the explainability vector, processing the first set of features to generate a second set of features such that each feature in the second set of features has a correlation with the output of the upstream machine learning model that is above a correlation threshold” (Cheng [0062] “The evaluation engine 140 can additionally process the model predictions and the model explanations, e.g., to compute cumulative SHAP values […] In some examples, the evaluation engine 140 can automatically select the top feature attributions that explain some predetermined threshold, e.g., 80%, of the model prediction”);
It would have been obvious to one having ordinary skill in the art at the time that the invention was effectively filed to combine the teachings of Bacco and Kokalj with that of Cheng since a combination of known methods would yield predictable results. As shown in Cheng, only top correlations are kept. This allows for better explanations to be selected and therefore would improve the classification and explanations of other models.
Bacco further teaches “selecting a third set of features for use by a downstream machine learning model, wherein the third set of features represents input text data” (figure 1 transformer 2 has input text data);
“based on the third set of features and the explanative factor, generating a fourth set of features” (the combined data above represents the fourth set of features);
“training the downstream machine learning model to output sentiment classifications, wherein the downstream machine learning model uses the fourth set of features as input” (pg. 7 §3.2.1 “After T1 has elaborated the N sequences, the new generated representations {r1...rN} are stacked together to become the input of T2. T2 then outputs a contextual representation ci for the i-th sentence that depends on the other sentences (ci = f(r1...rN)).”); and
“using the downstream machine learning model, generating the sentiment classifications for input text data represented by the third set of features” (pg. 10 §5 “The proposed models were evaluated for both sentiment analysis and explainability outcomes.” i.e. sentiment classification).
Note that independent claims 2 and 12 recite the same substantial limitations as independent claim 1, only differing in embodiments and being slightly broader. There differences in embodiments are obvious variations of another and therefore the claims are subject to the same rejection.
Regarding claims 5 and 15, the Bacco, Kokalj, and Cheng references have been addressed above. Bacco further teaches “wherein processing the second set of features and the output of the upstream machine learning model to generate the explanative factor comprises: generating an encoding map which translates the first set of features to the second set of features” (figure 1 shows the features after transformer 1 being translated or processed);
“using the output of the upstream machine learning model and the explainability vector, generating an embedding vector” (pg. 7 §3.2.1 “The input of the first transformer is a sequence of t tokens, while the output is an embedding representation of that sequence”); and
“based on the encoding map and the embedding vector, generating the explanative factor” (previous citation “After T1 has elaborated the N sequences, the new generated representations {r1...rN} are stacked together to become the input of T2. T2 then outputs a contextual representation ci for the i-th sentence that depends on the other sentences (ci = f(r1...rN)). By merging these contextual representations we obtain an unique document representation d = U(c1...cN”)
Regarding claims 8 and 18, the Bacco, Kokalj, and Cheng references have been addressed above. Kokalj further teaches “wherein: the upstream machine learning model is defined by a set of parameters comprising a matrix of weights for a supervised classifier algorithm” (Kokalj og. 1 §2 “BERT (Devlin et al., 2019) is a large pretrained language model based on the transformer neural network architecture (Vaswani et al., 2017). Nowa days, BERT models exist in many mono- and multilingual variants”); and
“the explainability vector is extracted from the set of parameters using a Local Interpretable Model-agnostic Explanations method” (Kokalj pg. 2 left col. “The visualization approaches implemented in the explanation methods LIME and SHAP are primarily designed for explanations of tabular data and images. Although the visualization with LIME includes adjustments for text data, the resulting explanations are presented in the form of histograms that are sometimes hard to understand, as Figure 1 shows”)
Claim(s) 6 and 16 is/are rejected under 35 U.S.C. 103 as being unpatentable over Bacco, Kokalj, and Cheng, further in view of Shih etal. US 2023/0162028.
Regarding claims 6 and 16, the Bacco, Kokalj, and Cheng references have been addressed above. They do not explicitly teach the claim limitations. Shih however teaches “wherein processing the first set of features to generate the second set of features comprises applying feature engineering using a multi-relational decision tree learning algorithm on the first set of features” (Shih [0110] “For GBM models, the decision trees themselves can be visualized, or common model explanation libraries such as SHAP (Shapley additive explanations) and LIME (local interpretable model-agnostic explanations) may be in use in many production systems. Since the final model benefiting from the feature engineering is chosen to be sufficient in interpretability, the same techniques can be applied. Therefore, the missing link in the interpretability evaluation may be the feature engineering itself. Since the feature engineering is based on spline (or other curve) fitting on the raw data, the spline itself can be used as an explanation medium on the data. The splines outline the general trends of the features as learned by the source model and can be examined independently.”)
It would have been obvious to one having ordinary skill in the art at the time that the inventions were effectively filed to combine the teachings of Bacco, Kokalj, and Cheng with that of Shih since as stated, “Since the feature engineering is based on spline (or other curve) fitting on the raw data, the spline itself can be used as an explanation medium on the data. The splines outline the general trends of the features as learned by the source model and can be examined independently.” Shih [0110]. This allows for better machine learning/classification.
Claim(s) 7 and 17 is/are rejected under 35 U.S.C. 103 as being unpatentable over Bacco, Kokalj, and Cheng, further in view of Dalalyan, Arnak S. "Exponential weights in multivariate regression and a low-rankness favoring prior." [herein Dala]
Regarding claims 7 and 17, the Bacco, Kokalj, and Cheng references have been addressed above. They do not explicitly teach the claim limitations. Dala however teaches “wherein: the upstream machine learning model is defined by a set of parameters comprising a matrix of weights for a multivariate regression algorithm” (Dala abstract “We establish theoretical guarantees for the expected prediction error of the exponentially weighted aggregate in the case of multivariate regression that is when the label vector is multidimensional.”); and
It would have been obvious to one having ordinary skill in the art at the time that the inventions were effectively filed to combine the teachings of Bacco, Kokalj, and Cheng with that of Dala since a combination of known methods would yield predictable results. As shown in Dala, multivariate regression models are known in the art. Therefore extracting explanations using other techniques as shown below would operate as expected.
Kokalj further teaches “the explainability vector is extracted from the set of parameters using a Shapley Additive Explanation method” (Kokalj pg. 2 right col. “Various approaches have been proposed to inter pret neural text classifiers. Some of them focus on adapting existing SHAP based explanation methods by improving different aspects, e.g., the word masking”)
Claim(s) 9 and 19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Bacco, Kokalj, and Cheng, further in view of Bordt, Sebastian, and Ulrike von Luxburg. "From shapley values to generalized additive models and back."
Regarding claims 9 and 19, the Bacco, Kokalj, and Cheng references have been addressed above. They do not explicitly teach the claim limitations. Bordt however teaches “wherein: the upstream machine learning model is defined by a set of parameters comprising a vector of coefficients for a generalized additive model” (Bordt abstract “. This work offers a partial reconciliation between the two by establishing a correspondence between Shapley Values and Generalized Addi tive Models (GAMs)”); and
“the explainability vector is extracted from the vector of coefficients in the generalized additive model” (pg. 4 §4.1
PNG
media_image2.png
600
460
media_image2.png
Greyscale
)
It would have been obvious to one having ordinary skill in the art at the time that the inventions were effectively filed to combine the teachings of Bacco, Kokalj, and Cheng with that of Bordt since a combination of known methods would yield predictable results. As shown in Bordt, GAMs are known in the art, as well as understanding explainability from it. These are known techniques and when the specific GAM is applied, the explanation can be obtained using known methods.
Claim(s) 10 and 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Bacco, Kokalj, and Cheng, further in view of He et al. US 2021/0327563.
Regarding claims 10 and 20, the Bacco, Kokalj, and Cheng references have been addressed above. They do not explicitly teach the claim limitations. He however teaches “wherein: the upstream machine learning model is defined by a set of parameters comprising a matrix of weights for a convolutional neural network algorithm” (He [0084] “a machine learning model 42 is a classifying machine learning model or a predicting machine learning model including, but not limited to, (1) a deep neural network (e.g., a convolutional neural network, a recurrent neural network, etc.) and (2) a supervised learning machine (e.g., a linear or nonlinear support vector machine, a boosting classifier, etc.)”); and
“the explainability vector is extracted from the set of parameters using a Gradient Class Activation Mapping method” (He [0087] “Also of importance is a capability to execute any type of relevancy mapping for ascertaining an explanation of the prediction or the classification whereby such relevancy mapping may be used to generate salient image data 36a of the present disclosure (e.g., back propagation, guided back propagation, deconvolution, class activation mapping, gradient class activation mapping, etc.”)
It would have been obvious to one having ordinary skill in the art at the time that the inventions were effectively filed to combine the teachings of Bacco, Kokalj, and Cheng with that of He since a combination of known methods would yield predictable results. As shown in He, CNNs are known classifiers and using gradient class activation mapping is a known method to generate explanations. As these techniques are known and well-described, they would operate as expected when using a CNN as a classifier.
Claim(s) 11 is/are rejected under 35 U.S.C. 103 as being unpatentable over Bacco, Kokalj, and Cheng, further in view of Salazar, Sebastian, Samuel Denton, and Ansaf Salleb-Aouissi. "Counterfactual explanations for support vector machine models."
Regarding claim 11, the Bacco, Kokalj, and Cheng references have been addressed above. They do not explicitly teach the claim limitations. Salazar however teaches “wherein: the upstream machine learning model is defined by a set of parameters comprising a hyperplane matrix for a support vector machine algorithm” (Salazar §1 “Support Vector Machines (SVMs) are one of the best performing discriminative models in machine learning. Despite recent advancements in areas like Deep Learning, SVMs manage to remain competitive and come with several theoretical guarantees on stability and sample complexity. Support vector machines are particularly good for high dimensional data since they belong to a class of learners that learn hyperplanes with low ℓ2 norm.”); and
“the explainability vector is extracted from the set of parameters using a counterfactual explanation method” (abstract “we introduce two novel scale invariant cost functions for assessing the quality of counterfactual explanations and use them to evaluate the quality of our approach with a real medical dataset. Finally, we build a support vector machine model to predict whether law students will pass the Bar exam using protected features, and used our algorithms to uncover the inherent biases of the SVM”)
It would have been obvious to one having ordinary skill in the art at the time that the inventions were effectively filed to combine the teachings of Bacco, Kokalj, and Cheng with that of Salazar since a combination of known methods would yield predictable results. As shown in Salazar, SVMs are well-known in literature, and counterfactual explanations are a known technique to use in the context of SVMs. Therefore these known techniques operate together as expected when using an SVM as a model.
Allowable Subject Matter
Claims 3-4 and 13-14 are 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.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
Laktionov, Ivan, et al. "An explainable AI approach to agrotechnical monitoring and crop diseases prediction in Dnipro region of Ukraine." Journal of Artificial Intelligence and Soft Computing Research 13.4 (2023): 247-272.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to KEVIN W FIGUEROA whose telephone number is (571)272-4623. The examiner can normally be reached Monday-Friday, 10AM-6PM EST.
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, MIRANDA HUANG can be reached at (571)270-7092. 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.
KEVIN W FIGUEROA
Primary Examiner
Art Unit 2124
/Kevin W Figueroa/Primary Examiner, Art Unit 2124
Prosecution Insights