International Business Machines Corporation

Google Inc.

Vikas Sindhwani

Tara N Sainath

Jianying Hu

Jennifer Ceil Lai

Aleksandra Mojsilovic

Aurelie C Lozano

Kevin Singley

Shivakumar Vaithyanathan

Berthold Reinwald

Mrinal Kalakrishnan

Brian E D Kingsbury

Edison Lao Ting

Yuanyuan Tian

David Paul Woodruff

Rajasekar Krishnamurthy

Edwin Peter Dawson Pednault

Richard D Lawrence

Haim Avron

Douglas Ronald Burdick

Shirish Tatikonda

Prem Melville

Krzysztof Marcin Choromanski

Amol N Ghoting

Arindam Banerjee

Saha Ankan

Shiva P Kasiviswanathan

Bachir El Khadir

Jacob Varley

Atil Iscen

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for implementing long-short term memory layers with compressed gating functions. One of the systems includes a first long short-term memory (LSTM) layer, wherein the first LSTM layer is configured to, for each of the plurality of time steps, generate a new layer state and a new layer output by applying a plurality of gates to a current layer input, a current layer state, and a current layer output, each of the plurality of gates being configured to, for each of the plurality of time steps, generate a respective intermediate gate output vector by multiplying a gate input vector and a gate parameter matrix. The gate parameter matrix for at least one of the plurality of gates is a structured matrix or is defined by a compressed parameter matrix and a projection matrix.

Compressed recurrent neural network models

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for optimizing the determination of control policies for robots through the performance of simulations of robots and real-world context to determine control policy parameters.

Determining control policies for robots with noise-tolerant structured exploration

Techniques are described herein for generating a dynamical systems control policy. A non-parametric family of smooth maps is defined on which vector-field learning problems can be formulated and solved using convex optimization. In some implementations, techniques described herein address the problem of generating contracting vector fields for certifying stability of the dynamical systems arising in robotics applications, e.g., designing stable movement primitives. These learning problems may utilize a set of demonstration trajectories, one or more desired equilibria (e.g., a target point), and once or more statistics including at least an average velocity and average duration of the set of demonstration trajectories. The learned contracting vector fields may induce a contraction tube around a targeted trajectory for an end effector of the robot. In some implementations, the disclosed framework may use curl-free vector-valued Reproducing Kernel Hilbert Spaces.

Generating a robot control policy from demonstrations collected via kinesthetic teaching of a robot

Learning to effectively imitate human teleoperators, even in unseen, dynamic environments is a promising path to greater autonomy, enabling robots to steadily acquire complex skills from supervision. Various motion generation techniques are described herein that are rooted in contraction theory and sum-of-squares programming for learning a dynamical systems control policy in the form of a polynomial vector field from a given set of demonstrations. Notably, this vector field is provably optimal for the problem of minimizing imitation loss while providing certain continuous-time guarantees on the induced imitation behavior. Techniques herein generalize to new initial and goal poses of the robot and can adapt in real time to dynamic obstacles during execution, with convergence to teleoperator behavior within a well-defined safety tube.

Generating a robot control policy from demonstrations

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for a system configured to plan actions to be performed by a robotic agent interacting with an environment to accomplish an objective by determining an optimized trajectory of state&#x2014;action pairs for accomplishing the objective. The system maintains a current optimized trajectory and a current trust region radius, and optimizes a localized objective within the current trust region radius of the current optimized trajectory to determine a candidate updated optimized trajectory. The system determines whether the candidate updated optimized trajectory improves over the current optimized trajectory. In response to determining that the candidate updated optimized trajectory improves over the current optimized trajectory, the system updates the current optimized trajectory to the candidate updated optimized trajectory and updates the current trust region radius.

Selecting actions to be performed by a robotic agent

A system, method and computer program product for quickly and approximately solving structured regression problems. In one aspect, the system, method and computer program product are applied to problems that arise naturally in various statistical modeling settings&#x2014;when the design matrix is a Vandermonde matrix or a sequence of such matrices. Using the Vandermonde matrix structure further accelerates the solution of the regression problem, achieving running times that are faster than 'input sparsity'. The modeling framework speedup benefits of randomized regression for solving structured regression problems.

Sketching structured matrices in nonlinear regression problems

Systems and methods for reducing a number of training parameters in a deep belief network (DBN) are provided. A method for reducing a number of training parameters in a deep belief network (DBN) comprises determining a network architecture including a plurality of layers, using matrix factorization to represent a weight matrix of a final layer of the plurality of layers as a plurality of matrices, and training the DBN having the plurality of matrices.

Low-rank matrix factorization for deep belief network training with high-dimensional output targets

A method, system and computer program product for inferring topic evolution and emergence in a set of documents. In one embodiment, the method comprises forming a group of matrices using text in the documents, and analyzing these matrices to identify a first group of topics as evolving topics and a second group of topics as emerging topics. The matrices includes a first matrix X identifying a multitude of words in each of the documents, a second matrix W identifying a multitude of topics in each of the documents, and a third matrix H identifying a multitude of words for each of the multitude of topics. These matrices are analyzed to identify the evolving and emerging topics. In an embodiment, the documents form a streaming dataset, and two forms of temporal regularizers are used to help identify the evolving topics and the emerging topics in the streaming dataset.

Inferring emerging and evolving topics in streaming text

In response to issues of high dimensionality and sparsity in machine learning, it is proposed to use a multiple output regression modeling module that takes into account information on groups of related predictor features and groups of related regressions, both given as input, and outputs a regression model with selected feature groups. Optionally, the method can be employed as a component in methods of causal influence detection, which are applied on a time series training data set representing the time-evolving content generated by community members, output a model of causal relationships and a ranking of the members according to their influence.

Group variable selection in spatiotemporal modeling

Systems and methods for processing Machine Learning (ML) algorithms in a MapReduce environment are described. In one embodiment of a method, the method includes receiving a ML algorithm to be executed in the MapReduce environment. The method further includes parsing the ML algorithm into a plurality of statement blocks in a sequence, wherein each statement block comprises a plurality of basic operations (hops). The method also includes automatically determining an execution plan for each statement block, wherein at least one of the execution plans comprises one or more low-level operations (lops). The method further includes implementing the execution plans in the sequence of the plurality of the statement blocks.

Systems and methods for processing machine learning algorithms in a MapReduce environment

A computer model finds an originating community member or members, for instance amongst bloggers, scientific researchers, or phenomena in phenomenological systems. In one embodiment, non-explicit causal relationships may be inferred from comparing blogs as a whole. Influence relationships are derived from a weighted, directed graph output and sources of influence are ranked. This implementation is useful for a variety of applications.

Inferring influence and authority

In a computerized social network, expert and user chat sessions are stored and rated probabilistically. Later user requests for information are met with an expert ranking, based on a balance of similarities between expert profile and questions; similarity between expert profile and prior chat sessions, and dynamically updated chat session ratings. New sessions can be rated automatically with reference to keywords distilled from past sessions responsive to user ratings&#x2014;and based on session length.

Growing community of inventors

Hawthorne, NY, United States of America

Vikas Sindhwani