International Business Machines Corporation

Pranay Kumar Lohia

Diptikalyan Saha

Sameep Mehta

Manish Anand Bhide

Samiulla Zakir Hussain Shaikh

Aniya Aggarwal

Kuntal Dey

Himanshu Gupta

Deepak Vijaykeerthy

Nitin Kumar Gupta

Kalapriya Kannan

Kush Raj Varshney

Rema Ananthanarayanan

Gagandeep Singh

Madhavi Katari

Naveen Panwar

Nishtha Madaan

Rajmohan Chandrahasan

Sandeep Hans

Ruchir Puri

Vijay Arya

Priyanka Agrawal

Karthikeyan Natesan Ramamurthy

Parag Jain

Atul S Kumar

Hima Patel

Anush Sankaran

Anirban Laha

Disha Shrivastava

Samuel Chung Hoffman

Shanmukha Chaitanya Guttula

Saket Gurukar

Rishabh Gupta

Rachamalla Anirudh Reddy

A computer-implemented method according to one embodiment includes identifying an occurrence of accuracy drift by a trained model; identifying data associated with the accuracy drift, utilizing a drift detection model (DDM) constructed for the trained model; applying the data associated with the accuracy drift to a decision tree to determine a feature space and specific subset of the data causing the accuracy drift; analyzing a distribution of features within the feature space for the specific subset of the data causing the accuracy drift to determine specific features of the data causing the accuracy drift; and returning the specific features of the data causing the accuracy drift.

Explaining accuracy drift in production data

Methods, systems, and computer program products for detecting contextual bias in text are provided herein. A computer-implemented method includes identifying, by a machine learning network, a protected attribute in one or more data samples; processing the identified data samples using a first sub-network of the machine learning network, wherein the first sub-network is configured to determine a plurality of contexts of the protected attribute across the identified data samples; determining an impact of each of the plurality of contexts on a second sub-network of the machine learning network, wherein the second sub-network of the machine learning network is configured to classify a given data sample into one of a plurality of classes; and adjusting the second sub-network of the machine learning to account for the impact of at least one of the plurality of contexts on the second sub-network.

Contextual impact adjustment for machine learning models

A computer implemented method transforms data. Responsive to receiving a data transformation of an input string to an output string, a computer system identifies mappable tokens in the input string that are mappable to the output string. The computer system creates a set of initial mappings for a set of common tokens in the mappable tokens. The set of initial mappings maps the set of common tokens from the input string to the output string. The computer system creates a set of user mappings that maps the mappable tokens from input string to the output string using a user input to the set of initial mappings. The computer system generates program code that transform input strings to output strings using the set of user mappings that maps the mappable tokens from input string to the output string, wherein the program code is used to transform input strings to output strings.

Data transformation methodology using generated program code and token mappings

Methods, systems, and computer program products for automatically modifying responses from generative models using artificial intelligence techniques are provided herein. A computer-implemented method includes obtaining data pertaining to at least one conversation involving at least one automated conversation exchange software program and at least one user; identifying, among words proposed by the at least one automated conversation exchange software program in connection with the at least one conversation, words qualifying as belonging to one or more predetermined categories by processing the obtained data using artificial intelligence techniques; determining, by processing the identified words and at least one word-based data source, one or more alternate words; modifying at least a portion of the proposed words by replacing at least a portion of the identified words with at least a portion of the one or more alternate words; and performing at least one automated action based on the modifying.

Automatically modifying responses from generative models using artificial intelligence techniques

A post-processing method, system, and computer program product for post-hoc improvement of instance-level and group-level prediction metrics, including training a bias detector that learns to detect a sample that has an individual bias greater than a predetermined individual bias threshold value with constraints on a group bias, applying the bias detector on a run-time sample to select a biased sample in the run-time sample having a bias greater than the predetermined individual bias threshold bias value, and suggesting a de-biased prediction for the biased sample.

Post-hoc improvement of instance-level and group-level prediction metrics

Methods, systems, and computer program products for determining data representative of bias within a model are provided herein. A computer-implemented method includes obtaining a first dataset on which a model was trained, wherein the first dataset contains protected attributes, and a second dataset on which the model was trained, wherein the protected attributes have been removed from the second dataset; identifying, for each of the one or more protected attributes in the first dataset, one or more attributes in the second dataset correlated therewith; determining bias among at least a portion of the identified correlated attributes; and outputting, to at least one user, identifying information pertaining to the one or more instances of bias.

Determining data representative of bias within a model

One embodiment provides a method, including: receiving a dataset and a model corresponding to a bias checker, wherein the bias checker detects bias within both the dataset and the model, based upon a bias checking algorithm and a bias checking policy, wherein the dataset comprises a plurality of attributes; testing the bias checking algorithm of the bias checker by (i) generating test cases that modify the dataset by introducing bias therein and (ii) running the bias checker against the modified dataset; testing the bias checking policy of the bias checker by generating a plurality of test cases and running the bias checker against the plurality of test cases; and providing a notification to a user regarding whether the bias checker failed to indicate bias for one or more of the plurality of attributes.

Testing bias checkers

One embodiment provides a method, including: receiving a target unstructured document for determining whether the target unstructured document comprises biased information; identifying an objective of the target unstructured document by extracting, from the target unstructured document, (i) entities and (ii) relationships between the entities; creating a structured knowledge base, wherein the creating comprises (i) creating an entry in the structured knowledge base corresponding to the target unstructured document, (ii) identifying other unstructured documents having a similarity to the target unstructured document, and (iii) generating an entry in the structured knowledge base corresponding to each of the other unstructured documents; applying a bias detection technique on the structured knowledge base; and providing an indication of whether the target unstructured document comprises bias.

Bias detection for unstructured text

A method, computer system, and a computer program product for generating deep learning model insights using provenance data is provided. Embodiments of the present invention may include collecting provenance data. Embodiments of the present invention may include generating model insights based on the collected provenance data. Embodiments of the present invention may include generating a training model based on the generated model insights. Embodiments of the present invention may include reducing the training model size. Embodiments of the present invention may include creating a final trained model.

Deep learning model insights using provenance data

Methods, systems, and computer program products for generating explanations for a semantic parser are provided herein. A computer-implemented method includes providing to a generative model (i) at least one query and (ii) a context of at least one dataset applicable to the at least one query, wherein the generative model generates a plurality of perturbations for the at least one input query based on the context; providing the plurality of perturbations as inputs to a context aware sequence-to-sequence model, thereby obtaining a plurality of outputs; and generating, for (i) an additional query provided as input to the context aware sequence-to-sequence model and (ii) a context applicable to the additional query, an explanation indicative of one or more parts of the additional query that contributes to an output corresponding to the additional query, based at least in part on the plurality of outputs corresponding to the perturbations.

Generating explanations for context aware sequence-to-sequence models

A method for blockchain certification of artificial intelligence factsheets that includes receiving by a computing device, an artificial intelligence model. The computing device generates an artificial intelligence factsheet based upon logic of the artificial intelligence model. The computing device generates a blockchain link for a blockchain. The blockchain link certifies the artificial intelligence factsheet. The computing device transmits the blockchain link certifying the artificial intelligence factsheet to other computing devices.

Artificial intelligence certification of factsheets using blockchain

A source of bias identification (SoBI) tool is provided that identifies sources of bias in a dataset. A bias detection operation is performed on results of a computer model, based on an input dataset, to generate groupings of values for a protected attribute corresponding to a detected bias in the operation of the computer model. The SoBI tool generates a plurality of sub-groups for each grouping of values. Each sub-group comprises an individual value, or a sub-range, for the protected attribute. The SoBI tool analyzes each of the sub-groups in the plurality of sub-groups, based on at least one source of bias identification criterion, to identify one or more sources of bias in the input dataset. The SoBI tool outputs a bias notification to an authorized computing device specifying the one or more sources of bias in the input dataset.

Bias source identification and de-biasing of a dataset

Methods, systems, and computer program products for improving trustworthiness of artificial intelligence models in presence of anomalous data are provided herein. A method includes obtaining a machine learning model and a set of training data; determining one or more anomalous data points in said set of training data; for a given one of said anomalous data points, identifying attributes that decrease confidence with respect to at least one output of said machine learning model; determining that a root cause of said decreased confidence corresponds to one of: a class imbalance issue related to said at least one attribute, a confused class issue related to said at least one attribute, a low density issue related to said at least one attribute, and an adversarial issue related to said at least one attribute; and performing step(s) to improve said confidence based at least in part on said determined root cause.

Trustworthiness of artificial intelligence models in presence of anomalous data

Methods, systems and computer program products for automated test case generation are provided herein. A computer-implemented method includes selecting sample input data as a test case for a system under test, executing the test case on the system under test to obtain a result, and applying the result to a local explainer function to obtain at least a portion of a corresponding decision tree. The method further includes determining at least one path constraint from the decision tree, solving the path constraint to obtain a solution, and generating at least one other test case for the system under test based at least in part on the solution of the path constraint. The steps of the method are illustratively repeated in each of one or more additional iterations until at least one designated stopping criterion is met. The resulting test cases form a test suite for testing of a deep neural network (DNN) or other system.

Automated test case generation for deep neural networks and other model-based artificial intelligence systems

Methods, systems, and computer program products for domain aware explainable anomaly and drift detection for multi-variate raw data using a constraint repository are provided herein. A computer-implemented method includes obtaining a set of data and information indicative of a domain of said set of data; obtaining constraints from a domain-indexed constraint repository based on said set of data and said information, wherein the domain-indexed constraint repository comprises a knowledge graph having a plurality of nodes, wherein each node comprises an attribute associated with at least one of a plurality of domains and constraints corresponding to the attribute; detecting anomalies in said set of data based on whether portions of said set of data violate said retrieved constraints; generating an explanation corresponding to each of the anomalies that describe the attributes corresponding to the violated constraints; and outputting an indication of the anomalies and the corresponding explanation.

Growing community of inventors

Bhagalpur, India

Pranay Kumar Lohia