Google Inc.

Abhinav Shrivastava

Saurabh Singh

Johannes Balle

George Dan Toderici

Nicholas M Johnston

Kevin Patrick Murphy

Sami Ahmad Abu-El-Haija

Deniz Oktay

Carl Martin Vondrick

Chen Sun

Rahul Sukthankar

Cordelia Luise Schmid

Sergio Guadarrama Cotado

Alireza Fathi

Johannes Ballé

Example aspects of the present disclosure are directed to systems and methods that learn a compressed representation of a machine-learned model (e.g., neural network) via representation of the model parameters within a reparameterization space during training of the model. In particular, the present disclosure describes an end-to-end model weight compression approach that employs a latent-variable data compression method. The model parameters (e.g., weights and biases) are represented in a 'latent' or “reparameterization” space, amounting to a reparameterization. In some implementations, this space can be equipped with a learned probability model, which is used first to impose an entropy penalty on the parameter representation during training, and second to compress the representation using arithmetic coding after training. The proposed approach can thus maximize accuracy and model compressibility jointly, in an end-to-end fashion, with the rate-error trade-off specified by a hyperparameter.

Compression of machine-learned models via entropy penalized weight reparameterization

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for receiving, by a neural network (NN), a dataset for generating features from the dataset. A first set of features is computed from the dataset using at least a feature layer of the NN. The first set of features i) is characterized by a measure of informativeness; and ii) is computed such that a size of the first set of features is compressible into a second set of features that is smaller in size than the first set of features and that has a same measure of informativeness as the measure of informativeness of the first set of features. The second set of features if generated from the first set of features using a compression method that compresses the first set of features to generate the second set of features.

Learning compressible features

Example aspects of the present disclosure are directed to systems and methods that learn a compressed representation of a machine-learned model (e.g., neural network) via representation of the model parameters within a reparameterization space during training of the model. In particular, the present disclosure describes an end-to-end model weight compression approach that employs a latent-variable data compression method. The model parameters (e.g., weights and biases) are represented in a 'latent' or 'reparameterization' space, amounting to a reparameterization. In some implementations, this space can be equipped with a learned probability model, which is used first to impose an entropy penalty on the parameter representation during training, and second to compress the representation using arithmetic coding after training. The proposed approach can thus maximize accuracy and model compressibility jointly, in an end-to-end fashion, with the rate-error trade-off specified by a hyperparameter.

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing visual tracking. In one aspect, a method comprises receiving: (i) one or more reference video frames, (ii) respective reference labels for each of a plurality of reference pixels in the reference video frames, and (iii) a target video frame. The reference video frames and the target video frame are processed using a colorization machine learning model to generate respective pixel similarity measures between each of (i) a plurality of target pixels in the target video frame, and (ii) the reference pixels in the reference video frames. A respective target label is determined for each target pixel in the target video frame, comprising: combining (i) the reference labels for the reference pixels in the reference video frames, and (ii) the pixel similarity measures.

Visual tracking by colorization

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing action localization in images and videos. In one aspect, a system comprises a data processing apparatus; a memory in data communication with the data processing apparatus and storing instructions that cause the data processing apparatus to perform image processing and video processing operations comprising: receiving an input comprising an image depicting a person; identifying a plurality of context positions from the image; determining respective feature representations of each of the context positions; providing a feature representation of the person and the feature representations of each of the context positions to a context neural network to obtain relational features, wherein the relational features represent relationships between the person and the context positions; and determining an action performed by the person using the feature representation of the person and the relational features.

Action localization in images and videos using relational features

Growing community of inventors

Mountain View, CA, United States of America

Abhinav Shrivastava