International Business Machines Corporation

Naveen Panwar

Anush Sankaran

Srikanth Govindaraj Tamilselvam

Shreya Khare

Diptikalyan Saha

Senthil Kumar Kumarasamy Mani

Nishtha Madaan

Deepak Vijaykeerthy

Rahul Aralikatte

Pranay Kumar Lohia

Tarun Tater

Kuntal Dey

Sameep Mehta

Neelamadhav Gantayat

Samiulla Zakir Hussain Shaikh

Hima Patel

Jassimran Kaur

Rahul A R

Senthil K K Mani

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

One embodiment provides a method, comprising: receiving an input sentence for a classification by a machine-learning model, where the classification is based upon a sentiment of the input sentence; splitting the input sentence into a plurality of tokens, each of the plurality of tokens corresponding to a term within the input sentence; creating a causal subgraph from the plurality of tokens, wherein the creating is based upon a causal relationship identified between tokens of the plurality of tokens; identifying, using the causal subgraph, tokens of the plurality of tokens influencing the classification; and generating, based upon the tokens of the plurality of tokens, a causal explanation for the classification, wherein the causal explanation identifies at least one portion of the input sentence resulting in the classification.

Generation of causal explanations for text models

Embodiments are disclosed for a method. The method includes receiving an annotation set for a machine learning model. The annotation set includes multiple data points relevant to a task for the machine learning model. The method also includes determining total weights corresponding to the data points. The total weights are determined based on multiple ordering constraints indicating multiple data classes and corresponding weights. The corresponding weights represent a relative priority of the data classes with respect to each other. The method further includes generating an ordered annotation set from the annotation set. The ordered annotation set includes the data points in a sequence based on the determined total weights.

Ordering annotation sets for machine learning

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

Methods, systems, and computer program products for multi-modal construction of deep learning networks are provided herein. A computer-implemented method includes extracting, from user-provided multi-modal inputs, one or more items related to generating a deep learning network; generating a deep learning network model, wherein the generating includes inferring multiple details attributed to the deep learning network model based on the one or more extracted items; creating an intermediate representation based on the deep learning network model, wherein the intermediate representation includes (i) one or more items of data pertaining to the deep learning network model and (ii) one or more design details attributed to the deep learning network model; automatically converting the intermediate representation into source code; and outputting the source code to at least one user.

Multi-modal construction of deep learning networks

A method, computer system, and a computer program product for debugging a deep neural network is provided. The present invention may include identifying, automatically, one or more debug layers associated with a deep learning (DL) model design/code, wherein the identified one or more debug layers include one or more errors, wherein a reverse operation is introduced for the identified one or more debug layers. The present invention may then include presenting, to a user, a debug output based on at least one break condition, wherein in response to determining the at least one break condition is satisfied, triggering the debug output to be presented to the user, wherein the presented debug output includes a fix for the identified one or more debug layers in the DL model design/code and at least one actionable insight.

Debugging deep neural networks

One embodiment provides a method, including: mining a plurality of deep-learning models from a plurality of input sources; extracting information from each of the deep-learning models, by parsing at least one of (i) code corresponding to the deep-learning model and (ii) text corresponding to the deep-learning model; identifying, for each of the deep-learning models, operators that perform operations within the deep-learning model; producing, for each of the deep-learning models and from (i) the extracted information and (ii) the identified operators, an ontology comprising terms and features of the deep-learning model, wherein the producing comprises populating a pre-defined ontology format with features of each deep-learning model; and generating a deep-learning model catalog comprising the plurality of deep-learning models, wherein the catalog comprises, for each of the deep-learning models, the ontology corresponding to the deep-learning model.

Deep-learning model catalog creation

Methods, systems, and computer program products for combined data pre-process and architecture search for deep learning models are provided herein. A computer-implemented method includes obtaining data corresponding to a deep learning task; performing, based on the deep learning task and the data, a multi-objective learning process to select an optimal combination of (i) a deep learning architecture for the deep learning task and (ii) a data pre-processing strategy to be applied to the data, the data pre-processing strategy comprising one or more pre-processing steps; pre-processing the data for the selected deep learning architecture based on the data pre-processing strategy; and providing the pre-processed data as input to the selected deep learning architecture to perform the deep learning task.

Combined data pre-process and architecture search for deep learning models

Techniques for the suggestion and completion of deep learning models are disclosed including receiving a set of data and determining at least one property of the data. A plurality of characteristics of a computing device and a plurality of deep learning models are received and a score for each of the plurality of deep learning models is determined based on the received computing device characteristics and the determined at least one property of the data. The plurality of deep learning models are ranked for presentation to a user based on the determined scores. One or more of the deep learning models are presented on a display based on the ranking. A selection of one of the deep learning models is received and the selected deep learning model is trained using the set of data.

Suggestion and completion of deep learning models from a catalog

One embodiment provides a method, comprising: identifying, using a processor, an individual in a video segment; ascertaining, from at least the video segment, a viewpoint expressed by the individual on a topic, wherein the viewpoint comprises at least one of: a stance of the individual and a sentiment of the individual toward the topic; identifying, using a processor, a superset video comprising the video segment, wherein the superset video is an originally published video; ascertaining, based at least in part on the superset video, an overarching viewpoint of the individual on the topic; determining whether an inconsistency exists between the viewpoint expressed by the individual in the video segment and the overarching viewpoint of the individual ascertained in the superset video; and alerting, responsive to determining that an inconsistency exists, a user that the video segment contains the inconsistency.

Growing community of inventors

Bengaluru, India

Naveen Panwar