Google Inc.

Alexander Mordvintsev

Matthew Nirvan Sharifi

Jakob Foerster

Abhanshu Sharma

Philip Haeusser

A computing device is disclosed that includes at least one processor and a machine-learned model. The machine-learned model can include a plurality of blocks and one or more residual connections between two or more of the plurality of blocks. The machine-learned model can be configured to receive a model input and, in response to receipt of the model input, output a model output. The machine-learned model can be configured to perform operations including determining a resource allocation parameter that corresponds to a desired allocation of system resources to the machine-learned model at an inference time: deactivating a subset of the plurality of blocks of the machine-learned model based on the resource allocation parameter: inputting the model input into the machine-learned model with the subset of the plurality of blocks deactivated; and receiving, as an output of the machine-learned model, the model output.

Systems and methods for providing a machine-learned model with adjustable computational demand

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a neural network. One of the methods includes obtaining a batch of labeled training items and a batch of unlabeled training items; processing the labeled training items and the unlabeled training items using the neural network and in accordance with current values of the network parameters to generate respective embeddings; determining a plurality of similarity values, each similarity value measuring a similarity between the embedding for a respective labeled training item and the embedding for a respective unlabeled training item; determining a respective roundtrip path probability for each of a plurality of roundtrip paths; and performing an iteration of a neural network training procedure to determine a first value update to the current values of the network parameters that decreases roundtrip path probabilities for incorrect roundtrip paths.

Semi-supervised training of neural networks

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for hiding information using neural networks. One of the methods includes maintaining data mapping each of a plurality of classes to a respective piece of information that may potentially be hidden in a received data item; receiving a new data item; receiving data identifying a first piece of information to be hidden in the new data item; and modifying the new data item to generate a modified data item that, when processed by a neural network configured to classify input data items belonging to one of the plurality of classes, is classified by the neural network as belonging to a first class of the plurality of classes that is mapped to the first piece of information in the maintained data.

Hiding and detecting information using neural networks

Example aspects of the present disclosure are directed to systems and methods that perform image style transfer for three-dimensional models. In some implementations, the systems and methods can use machine-learned models such as, for example, convolutional neural networks to generate image style and content information used to perform style transfer. The systems and methods of the present disclosure can operate in a rendered image space. In particular, a computing system can iteratively modify an attribute rendering map (e.g., texture map, bump map, etc.) based on information collected from a different rendering of the model at each of a plurality of iterations, with the end result being that the attribute rendering map mimics the style of one or more reference images in content-preserving way. In some implementations, a computation of style loss at each iteration can be performed using multi-viewpoint averaged scene statistics, instead of treating each viewpoint independently.

Image style transfer for three-dimensional models

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for image generation using neural networks. In one of the methods, an initial image is received. Data defining an objective function is received, and the objective function is dependent on processing of a neural network trained to identify features of an image. The initial image is modified to generate a modified image by iteratively performing the following: a current version of the initial image is processed using the neural network to generate a current objective score for the current version of the initial image using the objective function; and the current version of the initial image is modified to increase the current objective score by enhancing a feature detected by the processing.

Image generation using neural networks

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for correcting a corrupted data sample using a trained deep neural network, the method including obtaining a feature representation of a corrupted data sample; and modifying the feature representation of the corrupted data sample to generate a feature representation of a corrected data sample by iteratively processing a current version of the feature representation of the corrupted data sample using the trained deep neural network to generate a current corruption score for the current version of the feature representation of the corrupted data sample and generating a less-corrupted version of the feature representation by performing an iteration of gradient descent against the current version of the feature representation of the corrupted data sample to reduce the current corruption score.

Growing community of inventors

Zurich, Switzerland

Alexander Mordvintsev