International Business Machines Corporation

Columbia University

Michele Merler

John Richard Smith

Rosario A Uceda-Sosa

Hui Wu

Dhiraj Joshi

Rogerio Schmidt Feris

Bishwaranjan Bhattacharjee

Stephen Carl Hammer

Liangliang Cao

Rameswar Panda

John Joseph Kent

Quoc-Bao Nguyen

Mudhakar Srivatsa

Wei Zhang

Raghu Kiran Ganti

Alfio Massimiliano Gliozzo

Ulrich Alfons Finkler

Nirmit V Desai

Jae-Eun Park

John Ronald Kender

Avirup Sil

Jae-Wook Ahn

Taesun Moon

Cicero Nogueira Dos Santos

Shreeranjani Srirangamsridharan

Mauro Martino

Mayoore Selvarasa Jaiswal

Sebastien Gilbert

Dean Steuer

Yang Zhao

Aashka Trivedi

Apurv Gupta

Shyama Prosad Chowdhury

Chidansh Amitkumar Bhatt

Graeme William Blackwood

Systems, computer-implemented methods, and computer program products to facilitate improving a mono-lingual language model using pre-trained models and parallel data are provided. According to an embodiment, a system can comprise a processor that executes components stored in memory. The computer executable components can comprise a training component that can use a first language model to improve a mono-lingual second language model through mask language modeling (MLM) training and contrastive learning. The computer executable components can comprise a data component that can collect parallel data for the first language model and the mono-lingual second language model. The computer executable components can comprise a tokenizing component that can tokenize the parallel data for the first language model and the mono-lingual second language model. The computer executable components can comprise an alignment component that can create an alignment of the tokenized parallel data.

Mono-lingual language models using parallel data

An embodiment includes identifying an initial plurality of sets of hyperparameter values at which to evaluate an objective function that relates hyperparameter values to performance values of a neural network. The embodiment also executes training processes on the neural network with the hyperparameters set to the each of the initial sets of hyperparameter values such that the training process provides an initial set of the performance values for the objective function. The embodiment also generates an approximation of the objective function using splines at selected performance values. The embodiment approximates a point at which the approximation of the objective function reaches a maximum value, then determines an updated set of hyperparameter values associated with the maximum value. The embodiment then executes a runtime process using the neural network with the hyperparameters set to the updated set of hyperparameter values.

Configuring a neural network using smoothing splines

A neural architecture search method, system, and computer program product that determines, by a computing device, a best fit language model of a plurality of language models that is a best fit for interpretation of a corpus of natural language and interprets, by the computing device, the corpus of natural language using the best fit language model.

Neural architecture search of language models using knowledge distillation

Described are techniques for oblique image rectification. The techniques include receiving an original image depicting an oblique view of a circular object and pre-processing the original image into an edge image. The techniques further include generating, by a machine learning model based on the edge image, a heatmap including an ellipse formed by the oblique view of the circular object. The techniques further include computing ellipse parameters describing the ellipse of the heatmap. The techniques further include performing, using the ellipse parameters, an affine transformation on the original image to generate a rectified image, where the rectified image converts the ellipse to a circle.

Oblique image rectification

Systems, methods and/or computer program products for controlling and visualizing topic modeling results using a topic modeling interface. The interface allows user directed exploration, understanding and control of topic modeling algorithms, while offering both semantic summaries and/or structure attribute explanations about results. Explanations and differentiations between results are presented using metrics such as cohesiveness and visual displays depicting hierarchical organization. Through user-manipulation of features of the interface, iterative changes are implemented through user-feedback, adjusting parameters, broadening or narrowing topic results, and/or reorganizing topics by splitting or merging topics. As users trigger visual changes to results being displayed, users can compare and contrast output from the topic modeling algorithm. With each change to parameters, users view different explanations informing the user why the changes being displayed occurred, providing users deeper understanding of the topic modeling process, how to manipulate parameters to achieve accurate topic results and adjust granularity of information presented.

Transparent and controllable topic modeling

A method and system for auto-curating a media are provided. Media content is received over the network interface. A set of markers is identified for the media content, each marker corresponding to one of a plurality of visible and audible cues in the media content. Segments in the media content are identified based on the identified set of markers. An excitement score is computed for each segment based on the identified markers that fall within the segment. A highlight clip is generated by identifying segments having excitement scores greater than a threshold.

Auto-curation and personalization of sports highlights

Techniques for content generation are provided. A plurality of discriminative terms is determined based at least in part on a first plurality of documents that are related to a first concept, and a plurality of positive exemplars and a plurality of negative exemplars are identified using the plurality of discriminative terms. A first machine learning (ML) model is trained to classify images into concepts, based on the plurality of positive exemplars and the plurality of negative exemplars. A second concept related to the first concept is then determined, based on the first ML model. A second ML model is trained to generate images based on the second concept, and a first image is generated using the second ML model. The first image is then refined using a style transfer ML model that was trained using a plurality of style images.

Automatic generation of content using multimedia

A method and system are provided. The method includes deriving a set of user attributes from an aggregate analysis of images and videos of a user. The deriving step includes recognizing, by a set of visual classifiers, semantic concepts in the images and videos of the user to generate visual classifier scores. The deriving step further includes deriving, by a statistical aggregator, the set of user attributes. The set of user attributes are derived by mapping the visual classifier scores to a taxonomy of semantic categories to be recognized in visual content. The deriving step also includes displaying, by an interactive user interface having a display, attribute profiles for the attributes and comparisons of the attribute profiles.

Individual and user group attributes discovery and comparison from social media visual content

A method includes utilizing two or more classifiers to calculate, for an input image, probability scores for a plurality of classes based on visual information extracted from the input image and semantic relationships in a classification hierarchy, wherein each of the two or more classifiers is associated with a given one of two or more levels in the classification hierarchy with each level in the classification hierarchy comprising a subset of the plurality of classes, and classifying the input image based on the calculated probability scores.

Image classification utilizing semantic relationships in a classification hierarchy

A method and systems are provided. A system includes a set of visual and textual classifiers for recognizing semantic concepts in a set of images and assigning semantic scores for the images to predict a gender of a user, and performing gender prediction from visual content and textual content in the images to respectively generate visual-based gender predictions and textual-based gender predictions. The system further includes a multimodal information fusion device for combining, using multimodal information fusion, the visual-based gender predictions, the textual-based gender predictions, and the semantic scores to infer a gender of a user.

Systems and methods for inferring gender by fusion of multimodal content

A method and systems are provided. The method includes recognizing semantic concepts in a set of images and assigning semantic scores for the images to predict a gender of a user. The method further includes performing gender prediction from visual content and textual content in the images to respectively generate visual-based gender predictions and textual-based gender predictions. The method also includes combining, using multimodal information fusion, the visual-based gender predictions, the textual-based gender predictions, and the semantic scores, to infer a gender of a user.

Techniques for tracking one or more objects at each position in an interval in a video input with the use of a Kalman filter, including obtaining a first location estimate of an object with an object detector, obtaining a second location estimate and a movement estimate of the object with an object tracker, determining a final estimate of a position and/or a velocity of the object with the Kalman filter.

Kalman filter approach to augment object tracking

Growing community of inventors

New York, NY, United States of America

Michele Merler