International Business Machines Corporation

David C Haws

Laxmi Priya Parida

Dan He

George Andrei Saon

Michael Alan Picheny

Samuel Jonathan Thomas

Dimitrios B Dimitriadis

Irina Rish

Brian E D Kingsbury

Xiaodong Cui

Zoltan Tueske

Mechanisms are provided for performing machine learning training of a computer model. A perturbation generator generates a modified training data comprising perturbations injected into original training data, where the perturbations cause a data corruption of the original training data. The modified training data is input into a prediction network of the computer model and processing the modified training data through the prediction network to generate a prediction output. Machine learning training is executed of the prediction network based on the prediction output and the original training data to generate a trained prediction network of a trained computer model. The trained computer model is deployed to an artificial intelligence computing system for performance of an inference operation.

Reducing exposure bias in machine learning training of sequence-to-sequence transducers

Various embodiments select markers for modeling epistasis effects. In one embodiment, a processor receives a set of genetic markers and a phenotype. A relevance score is determined with respect to the phenotype for each of the set of genetic markers. A threshold is set based on the relevance score of a genetic marker with a highest relevancy score. A relevance score is determined for at least one genetic marker in the set of genetic markers for at least one interaction between the at least one genetic marker and at least one other genetic marker in the set of genetic markers. The at least one interaction is added to a top-k feature set based on the relevance score of the at least one interaction satisfying the threshold.

Feature selection for efficient epistasis modeling for phenotype prediction

Audio features, such as perceptual linear prediction (PLP) features and time derivatives thereof, are extracted from frames of training audio data including speech by multiple speakers, and silence, such as by using linear discriminant analysis (LDA). The frames are clustered into k-means clusters using distance measures, such as Mahalanobis distance measures, of means and variances of the extracted audio features of the frames. A recurrent neural network (RNN) is trained on the extracted audio features of the frames and cluster identifiers of the k-means clusters into which the frames have been clustered. The RNN is applied to audio data to segment audio data into segments that each correspond to one of the cluster identifiers. Each segment can be assigned a label corresponding to one of the cluster identifiers. Speech recognition can be performed on the segments.

Using recurrent neural network for partitioning of audio data into segments that each correspond to a speech feature cluster identifier

Speaker diarization is performed on audio data including speech by a first speaker, speech by a second speaker, and silence. The speaker diarization includes segmenting the audio data using a long short-term memory (LSTM) recurrent neural network (RNN) to identify change points of the audio data that divide the audio data into segments. The speaker diarization includes assigning a label selected from a group of labels to each segment of the audio data using the LSTM RNN. The group of labels comprising includes labels corresponding to the first speaker, the second speaker, and the silence. Each change point is a transition from one of the first speaker, the second speaker, and the silence to a different one of the first speaker, the second speaker, and the silence. Speech recognition can be performed on the segments that each correspond to one of the first speaker and the second speaker.

Using long short-term memory recurrent neural network for speaker diarization segmentation

Various embodiments select features from a feature space. In one embodiment, a set of training samples and a set of test samples are received. The set of training samples includes a set of features and a class value. The set of test samples includes the set of features absent the class value. A relevancy with respect to the class value is determined for each of a plurality of unselected features based on the set of training samples. A redundancy with respect to one or more of the set of features is determined for each of the plurality of unselected features in the first set of features based on the set of training samples and the set of test samples. A set of features is selected from the plurality of unselected features based on the relevancy and the redundancy determined for each of the plurality of unselected features.

Transductive feature selection with maximum-relevancy and minimum-redundancy criteria

Various embodiments sort data. In one embodiment, a matrix D including a set of data values is received. A matrix Q is received, and includes a set of columns and a set of rows. The matrix Q further includes a sorting of each column of the matrix D. Each of these rows corresponds to a sorting. Each of a set of values in each of the set of columns in the matrix Q identifies a row in the matrix D. At least one sub-matrix D' of the matrix D is identified. A set of columns of the sub-matrix D&#x2032; is restricted to one or more columns of the matrix D. A processor sorts the sub-matrix D&#x2032; by rows based on the sorting of the set of columns of the matrix D as given in the matrix Q, and based on the set of data values in the matrix D.

Efficient sorting of large dimensional data

Various embodiments provide lossless compression of an enumeration space for genetic founder lines. In one embodiment, an input comprising a set of genetic founder lines and a maximum number of generations G is obtained. A set of genetic crossing templates of a height h is generated. A determination is made if at least a first genetic crossing template in the set of genetic crossing templates is redundant with respect to a second genetic crossing template in the set of genetic crossing templates. Based on the at least first genetic crossing template being redundant is redundant with respect to the second genetic crossing template, the at least first genetic crossing template is removed from the set of genetic crossing templates. This process of removing the at least first genetic crossing template from the set of genetic crossing templates the redundant creates an updated set of genetic crossing templates.

Growing community of inventors

New York, NY, United States of America

David C Haws