Google Inc.

Mohammad Norouzi

Quoc V Le

William Chan

Jeffrey Adgate Dean

Geoffrey Everest Hinton

Simon Kornblith

Ting Chen

Naveen Kumar

Jonathon Shlens

Samuel Bengio

Azalia Mirhoseini

David James Fleet

Chitwan Saharia

Rasmus Munk Larsen

Jonathan Ho

Tim Salimans

Hieu Hy Pham

Benoit Steiner

Sara Sabour Rouh Aghdam

Ofir Nachum

Dale Eric Schuurmans

Kelvin Xu

Yuefeng Zhou

Yonghui Wu

Samy Bengio

Gregory Sean Corrado

Karen Simonyan

Zhifeng Chen

Andrea Lynn Frome

Douglas Eck

Yoram Singer

Sander Etienne Lea Dieleman

Kevin Patrick Murphy

Michael Schuster

Tomas Mikolov

Kevin Jordan Swersky

Jesse Engel

Adam Roberts

David J Bieber

Sergio Guadarrama Cotado

Daniel Aaron Abolafia

Cinjon Resnick

Ryan Lienhart Dahl

Systems, methods, and computer program products for performing semi-supervised contrastive learning of visual representations are provided. For example, the present disclosure provides systems and methods that leverage particular data augmentation schemes and a learnable nonlinear transformation between the representation and the contrastive loss to provide improved visual representations. Further, the present disclosure also provides improvements for semi-supervised contrastive learning. For example, computer-implemented method may include performing semi-supervised contrastive learning based on a set of one or more unlabeled training data, generating an image classification model based on a portion of a plurality of layers in a projection head neural network used in performing the contrastive learning, performing fine-tuning of the image classification model based on a set of one or more labeled training data, and after performing the fine-tuning, distilling the image classification model to a student model comprising a relatively smaller number of parameters than the image classification model.

Systems and methods for contrastive learning of visual representations

A method for determining a placement for machine learning model operations across multiple hardware devices is described. The method includes receiving data specifying a machine learning model to be placed for distributed processing on multiple hardware devices; generating, from the data, a sequence of operation embeddings, each operation embedding in the sequence characterizing respective operations necessary to perform the processing of the machine learning model; processing the sequence of operation embeddings using a placement recurrent neural network in accordance with first values of a plurality network parameters of the placement recurrent neural network to generate a network output that defines a placement of the operations characterized by the operation embeddings in the sequence across the plurality of devices; and scheduling the machine learning model for processing by the multiple hardware devices by placing the operations on the multiple devices according to the placement defined by the network output.

Device placement optimization with reinforcement learning

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a neural network using a priority queue. One of the methods includes maintaining data identifying a set of K output sequences that were previously generated; selecting at least one of the output sequences from the set of output sequences; for each selected output sequence, determining a respective score; determining, for each selected sequence, a respective first update to the current values of the controller parameters; generating a batch of new output sequences using the controller neural network; obtaining a respective reward for each of the new output sequences; determining, from the new output sequences and the output sequences in the maintained data, the K output sequences that have the highest rewards; and modifying the maintained data.

Training neural networks using priority queues

A method includes receiving, by a computing device, training data comprising a plurality of pairs of images, wherein each pair comprises an image and at least one corresponding target version of the image. The method also includes training a neural network based on the training data to predict an enhanced version of an input image, wherein the training of the neural network comprises applying a forward Gaussian diffusion process that adds Gaussian noise to the at least one corresponding target version of each of the plurality of pairs of images to enable iterative denoising of the input image, wherein the iterative denoising is based on a reverse Markov chain associated with the forward Gaussian diffusion process. The method additionally includes outputting the trained neural network.

Image enhancement via iterative refinement based on machine learning models

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a sequence generation neural network. One of the methods includes obtaining a batch of training examples; for each of the training examples: processing the training network input in the training example using the neural network to generate an output sequence; for each particular output position in the output sequence: identifying a prefix that includes the system outputs at positions before the particular output position in the output sequence, for each possible system output in the vocabulary, determining a highest quality score that can be assigned to any candidate output sequence that includes the prefix followed by the possible system output, and determining an update to the current values of the network parameters that increases a likelihood that the neural network generates a system output at the position that has a high quality score.

Training sequence generation neural networks using quality scores

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a policy neural network used to select actions to be performed by a reinforcement learning agent interacting with an environment. In one aspect, a method includes obtaining path data defining a path through the environment traversed by the agent. A consistency error is determined for the path from a combined reward, first and last soft-max state values, and a path likelihood. A value update for the current values of the policy neural network parameters is determined from at least the consistency error. The value update is used to adjust the current values of the policy neural network parameters.

Training policy neural networks using path consistency learning

Provided are systems and methods for contrastive learning of visual representations. In particular, the present disclosure provides systems and methods that leverage particular data augmentation schemes and a learnable nonlinear transformation between the representation and the contrastive loss to provide improved visual representations. In contrast to certain existing techniques, the contrastive self-supervised learning algorithms described herein do not require specialized architectures or a memory bank. Some example implementations of the proposed approaches can be referred to as a simple framework for contrastive learning of representations or 'SimCLR.' Further example aspects are described below and provide the following benefits and insights.

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for neural machine translation. One of the systems includes an encoder neural network comprising: an input forward long short-term memory (LSTM) layer configured to process each input token in the input sequence in a forward order to generate a respective forward representation of each input token, an input backward LSTM layer configured to process each input token in a backward order to generate a respective backward representation of each input token and a plurality of hidden LSTM layers configured to process a respective combined representation of each of the input tokens in the forward order to generate a respective encoded representation of each of the input tokens; and a decoder subsystem configured to receive the respective encoded representations and to process the encoded representations to generate an output sequence.

Neural machine translation systems

Systems and methods for transforming grayscale images into color images using deep neural networks are described. One of the systems include one or more computers and one or more storage devices storing instructions that, when executed by one or more computers, cause the one or more computers to implement a coloring neural network, a refinement neural network, and a subsystem. The coloring neural network is configured to receive a first grayscale image having a first resolution and to process the first grayscale image to generate a first color image having a second resolution lower than the first resolution. The subsystem processes the first color image to generate a set of intermediate image outputs. The refinement neural network is configured to receive the set intermediate image outputs, and to process the set of intermediate image outputs to generate a second color image having a third resolution higher than the second resolution.

Transforming grayscale images into color images using deep neural networks

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for classifying data objects. One of the methods includes obtaining data that associates each term in a vocabulary of terms with a respective high-dimensional representation of the term; obtaining classification data for a data object, wherein the classification data includes a respective score for each of a plurality of categories, and wherein each of the categories is associated with a respective category label; computing an aggregate high-dimensional representation for the data object from high-dimensional representations for the category labels associated with the categories and the respective scores; identifying a first term in the vocabulary of terms having a high-dimensional representation that is closest to the aggregate high-dimensional representation; and selecting the first term as a category label for the data object.

Classifying data objects

Growing community of inventors

Sunnyvale, CA, United States of America

Mohammad Norouzi