Naver Corporation

Conduent Business Services, LLC

Xerox Corporation

Naver Labs Corporation

Julien Perez

Tomi Silander

Quentin Grail

Christopher R Dance

Guillaume M Bouchard

Claude Roux

Nicolas Monet

James Brinton Henderson

Arnaud Sors

Eric Jacques Guy Gaussier

Caroline Brun

Denys Proux

Diane Larlus-Larrondo

Adrian Corneliu Mos

William Radford

Michael Niemaz

Scott Peter Nowson

Mert Bulent Sariyildiz

Diana Nicoleta Popa

Michel Aractingi

Scott Nowson

Fei Liu

Theo Cachet

Seungsu Kim

David Emukpere

Bingbing Wu

Morgan Funtowicz

Systems and methods are disclosed for determining a policy to recommend transition in a position-representing space for a robotic device using a multi-critic architecture. To learn policy in a multi-critic architecture, a set of critics is defined pertaining to a position-representing space where each critic corresponds to a different objective function such as reach-reward, discovery-reward, and safety-reward. For each one of the critics of the set of critics, a learned value function in position-representing space is determined. The policy is learned based on the weighted feedback of the learned value functions to recommend transitions that are safe in the position-representing space. The multi-critic architecture minimizes interference between multiple reward functions and learns a safe and stable policy for the robotic device.

Systems and methods for skill learning with multiple critics

A computer-implemented method for performing few-shot imitation is disclosed. The method comprises obtaining at least one set of training data, wherein each set of training data is associated with a task and comprises (i) one of samples of rewards and a reward function, (ii) one of samples of state transitions and a transition distribution, and (iii) a set of first demonstrations, training a policy network embodied in an agent using reinforcement learning by inputting at least one set of first demonstrations of the at least one set of training data into the policy network, and by maximizing a risk measure or an average return over the at least one set of first demonstrations of the at least one set of training data based on respective one or more reward functions or respective samples of rewards, obtaining a set of second demonstrations associated with a new task, and inputting the set of second demonstrations and an observation of a state into the trained policy network for performing the new task.

Demonstration-conditioned reinforcement learning for few-shot imitation

A method of using a first neural network includes: by the first neural network, receiving a text; by the first neural network, receiving a question concerning the text; and by the first neural network, determining an answer to the question using the text, where the first neural network is trained to answer the question about the text adversarially by a second neural network that is trained to maximize a likelihood of failure of the first neural network to correctly answer questions.

Training of machine reading and comprehension systems

A system includes: a first module configured to, based on a set of target robot joint angles, generate a first estimated end effector pose and a first estimated latent variable that is a first intermediate variable between the set of target robot joint angles and the first estimated end effector pose; a second module configured to determine a set of estimated robot joint angles based on the first estimated latent variable and a target end effector pose; a third module configured to determine joint probabilities for the robot based on the first estimated latent variable and the target end effector pose; and a fourth module configured to, based on the set of estimated robot joint angles, determine a second estimated end effector pose and a second estimated latent variable that is a second intermediate variable between the set of estimated robot joint angles and the second estimated end effector pose.

Reachable manifold and inverse mapping training for robots

A confidence estimation system includes: a neural network including at least one an attention module including N heads configured to: generate attention matrices based on interactions between tokens for words in an input sequence of words, the input sequence of words including a word that is obscured; and determine the word that is obscured in the input sequence; and a confidence module configured to determine a confidence value indicative of a probability of the neural network correctly determining the word that is obscured, the confidence module determining the confidence value of the word that is obscured using a convolutional neural network that projects the attention matrices generated by the attention module over a multi-dimensional space, the attention matrices recording interactions between the tokens in the input sequence of words without information regarding the tokens for the words and the word that is obscured.

Self-attention-based confidence estimation of language models

A question answering system includes: a neural network tokenizer module configured to determine a token representation of a question to be answered and token representations of candidate paragraphs of a present reasoning path for the question, respectively; a neural network module configured to: transform the token representation of the question and the token representations of the candidate paragraphs of the present reasoning path into vector representations; and append a first variable to the vector representations to produce a second variable; a search module configured to: select the candidate paragraphs from a knowledge database to extend a present reasoning path based on lexical relevance of the candidate paragraphs to the question; and selectively add ones of the candidate paragraphs to the present reasoning path; and an answer inference network module configured to selectively determine an answer to the question based on multiple different portions of the present reasoning path.

Neural reasoning path retrieval for multi-hop text comprehension

A question answering system includes: a first encoder module configured to receive a question, the question including a first plurality of words, and encode the question into a first vector representation; a second encoder module configured to encode a document into a second vector representation, the document including a second plurality of words; a first reading module configured to generate a third vector representation based on the first and second vector representations; a first reformulation module configured to generate a first reformulated vector representation based on the first vector representation; a second reading module configured to generate a fifth vector representation based on the second vector representation and the first reformulated vector representation; a second reformulation module configured to generate a second reformulated vector representation based on first reformulated vector representation; and an answer module configured to determine an answer to the question based on the second reformulated vector representation.

Latent question reformulation and information accumulation for multi-hop machine reading

There is disclosed a computer implemented method that includes accessing a dataset having (1) a first set of questions including at least one pair of relational questions that correspond respectively with a pair of binary answers and (2) a second set of questions including at least another pair of relational questions that correspond respectively with a binary answer and a scalar answer. A question answering network is used to compute both a relational loss for the at least one pair of relational questions, and a relational loss for the at least another pair of relational questions. Both the relational loss for the at least one pair of relational questions and the relational loss for the at least another pair of relational questions are optimized, and a neural network model is trained with the optimized relational losses.

Training a question answering network using relational loss

A training system includes: a first training dataset including first entries, wherein each of the first entries includes: a first sentence; a second sentence; and an indicator of a relationship between the first and second sentences; a training module configured to: generate a second dataset including second entries based on the first entries, respectively, wherein each of the second entries includes: the first sentence of one of the first entries; the second sentence of the one of the first entries; a first surface realization corresponding to first facts regarding the first sentence; the indicator of the one of the first entries; and train a model using the second dataset and store the model in memory.

Attention over common-sense network for natural language inference

A method includes: initializing a list of token embeddings, each of the token embeddings corresponding to a tokenized word from text in a corpus of text; generating a graph for a group of consecutive words s from the text, said graph including binary relations between pairs of tokenized words of the group of consecutive words; selecting the token embeddings representing the words of the group of consecutive words from the list of token embeddings; computing a tensor of binary relations as the product between a matrix of the selected token embeddings and a tensor representing discourse relations, the computed tensor representing the binary relations between the pairs of tokenized words; computing a loss using the computed tensor; optimizing the list of token embeddings using the computed loss. The above may be repeated until the computed loss is within a predetermined range.

Implicit discourse relation classification with contextualized word representation

A computer implemented method for multi-document question answering is performed on a server communicating with a client device over a network. The method includes receiving a runtime question from a client device and retrieving runtime documents concerning the runtime question using a search engine. Runtime answer samples are identified in the retrieved runtime documents. A neural network model, trained using distant supervision and distance based ranking loss, is used to compute runtime scores from runtime question data representing the runtime question and from a runtime answer sample representing a first portion of text from the corpus of documents, where each runtime score represents a probability an answer to the runtime question is present in the runtime answer samples. A runtime answer is selected from the runtime answer samples corresponding to the highest runtime score sent to the client device.

Method and apparatus for multi-document question answering

A training system for training a trained model for use by a navigating robot to perform visual navigation includes memory including N base virtual training environments, each of the N base virtual training environments including a field of view at a location within an indoor space, where N is an integer greater than 1. A randomization module is configured to generate N varied virtual training environments based on the N base virtual training environments, respectively, by varying at least one characteristic of the respective N base virtual training environments. A training module is configured to train the trained model for use by the navigating robot to perform visual navigation based on a training set including: the N base virtual training environments; and the N varied virtual training environments.

Systems and methods for improving generalization in visual navigation

A method and system pre-trains a convolutional neural network for image recognition based upon masked language modeling by inputting, to the convolutional neural network, an image; outputting, from the convolutional neural network, a visual embedding tensor of visual embedding vectors; tokenizing a caption to create a list of tokens, at least one token having visual correspondence to the image received by the convolutional neural network; randomly selecting one of the tokens in the list of tokens to be masked, the selected token being taken as ground truth; computing, using a language model neural network, hidden representations of the tokens; using the hidden representation of the masked token, as a query vector, to pool the visual embedding vectors in the visual embedding tensor, attentively; predicting the masked token by mapping the pooled visual embedding vectors to the tokens; determining a prediction loss associated with the masked token; and back-propagating the prediction loss to the convolutional neural network to tune parameters thereof.

Method for training a convolutional neural network for image recognition using image-conditioned masked language modeling

A method and method for natural language generation employ a natural language generation model which has been trained to assign an utterance label to a new text sequence, based on features extracted from the text sequence, such as parts-of-speech. The model assigns an utterance label to the new text sequence, based on the extracted features. The utterance label is used to guide the generation of a natural language utterance, such as a question, from the new text sequence. The system and method find application in dialog systems for generating utterances, to be sent to a user, from brief descriptions of problems or solutions in a knowledge base.

Natural language generation, a hybrid sequence-to-sequence approach

A system and method for predicting a sequence of actions employ a Gated End-to-End Memory Policy Network (GMemN2NP), which includes a sequence of hop(s). Supporting memories of the hops include memory cells generated from observations made at different times. A sequence of actions is predicted, based on input agent-specific variables. For each action, the model, at each hop, outputs an updated controller state which is used as input to the next hop or, for the terminal hop, for computing the respective action. Each hop includes a transform gate mechanism which is used to control the influence of output of the supporting memories on the updated controller state. For the second and subsequent hops, respective actions are predicted, after using any intervening observations to update the supporting memories. The model is learned, on a training set of observations, to optimize the cumulative reward of a sequence of two or more actions.

Growing community of inventors

Grenoble, France

Julien Perez