International Business Machines Corporation

Massachusetts Institute of Technology

Rensselaer Polytechnic Institute

Shiyu Chang

Yang Zhang

Mo Yu

Wei Tan

Liana Liyow Fong

Chuang Gan

Ming Tan

Tommi S Jaakkola

Dakuo Wang

Haoyu Wang

Wang Zhou

Kaizhi Qian

Jinjun Xiong

Saloni Potdar

David Shing-Ki Kung

Sijia Liu

Payel Das

Yang Yu

David A Cox

David E Cox

Murray Scott Campbell

Xiaoxiao Guo

Norma Edith Sosa

Hui Su

Emrah Akin Sisbot

Li Zhang

Xiangyang Mou

Yupeng Gao

Minhao Cheng

Pin-yu Chen

Methods and apparatus for complementary evidence identification in natural language inference. A given question is obtained and a set of N passages is obtained from a database. A probability is determined, for each passage of the set of N passages, of a corresponding passage being a supportive passage for the given question and the set of N passages is ranked based on the determined probabilities. M passages that are ranked 1 to M of the set of N passages are selected. A set of L passages is selected based on a plurality of scores, each score assigned to a set of candidate passages of the set of N passages, each score being based on the determined probabilities, the selected M passages, and a weighted regulation parameter. The set of L passages is provided to a computerized machine learning system to answer the question based on the set of L passages.

Complementary evidence identification in natural language inference

To improve actual labels that are produced by a black box computer classifier system from inputs, identify, using an environment-aware predictor and an environment-agnostic predictor, a subset of the inputs. The subset of the inputs has a stable correlation with the actual labels across a plurality of environments. Identify the subset of the inputs as an explanatory rationale for the actual labels. Display the explanatory rationale with the actual labels to a consumer of the actual labels. Optionally, in response to the explanatory rationale failing a rubric established by the consumer, generate revised inputs by removing the explanatory rationale from the inputs; and produce revised labels by processing the revised inputs with the environment-agnostic predictor.

Game-theoretic invariant rationalization of machine-learning results

Techniques for generator model training are provided. A classifier model that was trained using one or more data samples in a target class is received, and a generative adversarial network (GAN) is trained to generate simulated data samples for the target class, comprising: generating a first simulated data sample using a generator model, computing a first discriminator loss by processing the first simulated data sample using a discriminator model, computing a classifier loss by processing the first simulated data sample using the classifier model, and refining the generator model based on the first discriminator loss and the classifier loss.

Generating data based on pre-trained models using generative adversarial models

A computer-implemented method is provided of using a machine learning model for disentanglement of prosody in spoken natural language. The method includes encoding, by a computing device, the spoken natural language to produce content code. The method further includes resampling, by the computing device without text transcriptions, the content code to obscure the prosody by applying an unsupervised technique to the machine learning model to generate prosody-obscured content code. The method additionally includes decoding, by the computing device, the prosody-obscured content code to synthesize speech indirectly based upon the content code.

Global prosody style transfer without text transcriptions

A method (and structure and computer product) to permit zero-shot voice conversion with non-parallel data includes receiving source speaker speech data as input data into a content encoder of a style transfer autoencoder system, the content encoder providing a source speaker disentanglement of the source speaker speech data by reducing speaker style information of the input source speech data while retaining content information and receiving target speaker input speech as input data into a target speaker encoder. The output of the content encoder and the target speaker encoder are combined in a decoder of the style transfer autoencoder, and the output of the decoder provides the content information of the input source speech data in a style of the target speaker speech information.

High-quality non-parallel many-to-many voice conversion

A current observation expressed in natural language is received. Entities in the current observation are extracted. A relevant historical observation is retrieved, which has at least one of the entities in common with the current observation. The current observation and the relevant historical observation are combined as observations. The observations and a template list specifying a list of verb phrases to be filled-in with at least some of the entities are input to a neural network, which can output the template list of the verb phrases filled-in with said at least some of the entities. The neural network can include attention mechanism. A reward associated with the neural network's output can be received and fed back to the neural network for retraining the neural network.

Extractive structured prediction in interactive systems

A method and system of determining an output label rationale are provided. A first generator receives a first class of data and selects one or more input features from the first class of data. A first predictor receives the one or more selected input features from the first generator and predicts a first output label. A second generator receives a second class of data and selects one or more input features from the second class of data. A second predictor receives the one or more selected input features from the second generator and predicts a second output label. A discriminator receives the first and second output labels and determines whether the selected one or more input features from the first class of data or the selected features of the one or more input features from the second class of data, more accurately represents the first output label.

Game-theoretic frameworks for deep neural network rationalization

A computer-implemented method includes using an embedding network to generate prototypical vectors. Each prototypical vector is based on a corresponding label associated with a first domain. The computer-implemented method also includes using the embedding network to generate an in-domain test vector based on at least one data sample from a particular label associated with the first domain and using the embedding network to generate an out-of-domain test vector based on at least one other data sample associated with a different domain. The computer-implemented method also includes comparing the prototypical vectors to the in-domain test vector to generate in-domain comparison values and comparing the prototypical vectors to the out-of-domain test vector to generate out-of-domain comparison values. The computer-implemented method also includes modifying, based on the in-domain comparison values and the out-of-domain comparison values, one or more parameters of the embedding network.

Out-of-domain encoder training

A method and system of training a natural language processing network are provided. A corpus of data is received and one or more input features selected therefrom by a generator network. The one or more selected input features from the generator network are received by a first predictor network and used to predict a first output label. A complement of the selected input features from the generator network are received by a second predictor network and used to predict a second output label.

Introspective extraction and complement control

Techniques regarding a zero-confidence adversarial attack are provided. For example, one or more embodiments described herein can comprise a system, which can comprise a memory that can store computer executable components. The system can also comprise a processor, operably coupled to the memory, and that can execute the computer executable components stored in the memory. The computer executable components can comprise an adversarial component that computes a perturbation that causes misclassification by a neural network classifier. The computer executable components can also comprise a restoration component that determines a normal vector to a constraint contour developed by the neural network classifier. Further, the computer executable components can comprise a projection component that determines a tangential vector to the constraint contour.

Margin based adversarial computer program

Techniques for training robust machine learning models for adversarial input data. Training data for a machine learning (ML) model is received. The training data includes a plurality of labels for data elements. First modified training data is generated by modifying one or more of the plurality of labels in the training data using parameterized label smoothing with a first optimization parameter. The ML model is trained using the first modified training data. The training includes updating a first one or more model weights in the ML model, and generating a second optimization parameter suitable for use in future parameterized label smoothing for future training of the ML model

Training robust machine learning models

A method, a structure, and a computer system for decomposing speech. The exemplary embodiments may include one or more encoders for generating one or more encodings of a speech input comprising rhythm information, pitch information, timbre information, and content information, and a decoder for decoding the one or more encodings.

Unsupervised speech decomposition

Techniques facilitating detection of conversation threads in unstructured channels are provided. A system can comprise a memory that stores computer executable components and a processor that executes the computer executable components stored in the memory. The computer executable components can comprise an extraction component that employs a model to detect conversation messages based on a defined confidence level and assigns the conversation messages to respective conversation thread categories. The computer executable components also can comprise a model component that trains the model on conversation messages that comprise respective text data, wherein the model is trained to detect the respective text data to the defined confidence level.

Facilitating detection of conversation threads in a messaging channel

Techniques that facilitate matrix factorization associated with graphics processing units are provided. In one example, a computer-implemented method is provided. The computer-implemented method can comprise loading, by a graphics processing unit operatively coupled to a processor, item features from a data matrix into a shared memory. The data matrix can be a matrix based on one or more user features and item features. The computer-implemented method can further comprise tiling and aggregating, by the graphics processing unit, outer products of the data matrix tiles to generate an aggregate value and approximating, by the graphics processing unit, an update to a user feature of the data matrix based on the aggregate value and the loaded item features.

Matrix factorization with approximate computing

Methods and systems perform incremental learning object detection in images and/or videos without catastrophic forgetting of previously-learned object classes. A two-stage neural network object detector is trained to locate and identify objects pertaining to an additional object class by iteratively updating the two-stage neural network object detector until an overall detection accuracy criterion is met. The updating is performed so as to balance minimizing a loss of an initial ability to locate and identify objects pertaining to the previously-learned object classes and maximizing an ability to additionally locate and identify the objects pertaining to the additional object class. Assessing whether the overall detection accuracy criterion is met compares outputs of an initial version of the two-stage neural network object detector with a current region proposal output by a current version of the two-stage neural network object detector to determining a region proposal distillation loss and a previously-learned-object identification distillation loss.

Growing community of inventors

Elmsford, NY, United States of America

Shiyu Chang