Waveone Inc.

Oren Rippel

Lubomir Dimitrov Bourdev

Sanjay Nair

Carissa Lew

Alexander Anderson

Steve Branson

Craig S Lytle

Hervé Guihot

Brandon Sprague

Kedar Tatwawadi

Steven Branson

An autoencoder is configured to encode content at different quality levels. The autoencoder includes an encoding system and a decoding system with neural network layers forming an encoder network and a decoder network. The encoder network and decoder network are configured to include branching paths through the networks that include different subnetworks. During deployment, content is provided to the encoding system with a quality signal indicating a quality at which the content can be reconstructed. The quality signal determines which of the paths through the encoder network are activated for encoding the content into one or more tensors, which are compressed into a bitstream and later used by the decoding system to reconstruct the content. The autoencoder is trained by randomly or systematically selecting different combinations of tensors to use to encode content and backpropagating error values from loss functions through the network paths associated with the selected tensors.

Dynamic control for a machine learning autoencoder

A deep learning based compression (DLBC) system trains multiple models that, when deployed, generates a compressed binary encoding of an input image that achieves a reconstruction quality and a target compression ratio. The applied models effectively identifies structures of an input image, quantizes the input image to a target bit precision, and compresses the binary code of the input image via adaptive arithmetic coding to a target codelength. During training, the DLBC system reconstructs the input image from the compressed binary encoding and determines the loss in quality from the encoding process. Thus, the models can be continually trained to, when applied to an input image, minimize the loss in reconstruction quality that arises due to the encoding process while also achieving the target compression ratio.

Deep learning based on image encoding and decoding

A compression system trains a compression model for an encoder and decoder. In one embodiment, the compression model includes a machine-learned in-loop flow predictor that generates a flow prediction from previously reconstructed frames. The machine-learned flow predictor is coupled to receive a set of previously reconstructed frames and output a flow prediction for a target frame that is an estimation of the flow for the target frame. In particular, since the flow prediction can be generated by the decoder using the set of previously reconstructed frames, the encoder may transmit a flow delta that indicates a difference between the flow prediction and the actual flow for the target frame, instead of transmitting the flow itself. In this manner, the encoder can transmit a significantly smaller number of bits to the receiver, improving computational efficiency.

Machine-learned in-loop predictor for video compression

A deep learning based compression (DLBC) system applies trained models to compress binary code of an input image to a target codelength. For a set of binary codes representing the quantized coefficents of an input image, the DLBC system applies a first model that is trained to predict feature probabilities based on the context of each bit of the binary codes. The DLBC system compresses the binary code via adaptive arithmetic coding based on the determined probability of each bit. The compressed binary code represents a balance between a reconstruction quality of a reconstruction of the input image and a target compression ratio of the compressed binary code.

Deep learning based adaptive arithmetic coding and codelength regularization

The compression system trains a machine-learned encoder and decoder through an autoencoder architecture. The encoder can be deployed by a sender system to encode content for transmission to a receiver system, and the decoder can be deployed by the receiver system to decode the encoded content and reconstruct the original content. The encoder is coupled to receive content and output a tensor as a compact representation of the content. The content may be, for example, images, videos, or text. The decoder is coupled to receive a tensor representing content and output a reconstructed version of the content. The compression system trains the autoencoder with a discriminator to reduce compression artifacts in the reconstructed content. The discriminator is coupled to receive one or more input content, and output a discrimination prediction that discriminates whether the input content is the original or reconstructed version of the content.

Using generative adversarial networks in compression

A machine learning (ML) task system trains a neural network model that learns a compressed representation of acquired data and performs a ML task using the compressed representation. The neural network model is trained to generate a compressed representation that balances the objectives of achieving a target codelength and achieving a high accuracy of the output of the performed ML task. During deployment, an encoder portion and a task portion of the neural network model are separately deployed. A first system acquires data, applies the encoder portion to generate a compressed representation, performs an encoding process to generate compressed codes, and transmits the compressed codes. A second system regenerates the compressed representation from the compressed codes and applies the task model to determine the output of a ML task.

Data compression for machine learning tasks

A deep learning based compression (DLBC) system generates a progressive representation of the encoded input image such that a client device that requires the encoded input image at a particular target bitrate can readily be transmitted the appropriately encoded data. More specifically, the DLBC system computes a representation that includes channels and bitplanes that are ordered based on importance. For a given target rate, the DLBC system truncates the representation according to a trained zero mask to generate the progressive representation. Transmitting a first portion of the progressive representation enables a client device with the lowest target bitrate to appropriately playback the content. Each subsequent portion of the progressive representation allows the client device to playback the content with improved quality.

Enhanced coding efficiency with progressive representation

An encoder system trains a compression model that includes an autoencoder model and a frame extractor model. The encoding portion of the autoencoder is coupled to receive a set of target frames and a previous state tensor for the set of target frames and generate compressed code. The decoding portion of the autoencoder is coupled to receive the compressed code and the previous state tensor for the set of frames and generate a next state tensor for the set of target frames. The frame extractor model is coupled to receive the next state tensor and generate a set of reconstructed frames that correspond to the set of target frames by performing one or more operations on the state tensor. The state tensor for the set of frames includes information from frames of the video that can be used by the frame extractor to generate reconstructed frames.

Machine-learning based video compression

A compression system includes an encoder and a decoder. The encoder can be deployed by a sender system to encode a tensor for transmission to a receiver system, and the decoder can be deployed by the receiver system to decode and reconstruct the encoded tensor. The encoder receives a tensor for compression. The encoder also receives a quantization mask and probability data associated with the tensor. Each element of the tensor is quantized using an alphabet size allocated to that element by the quantization mask data. The encoder compresses the tensor by entropy coding each element using the probability data and alphabet size associated with the element. The decoder receives the quantization mask data, the probability data, and the compressed tensor data. The quantization mask and probabilities are used to entropy decode and subsequently reconstruct the tensor.

Adaptive quantization

An enhanced encoder system generates residual bitstreams representing additional image information that can be used by an image enhancement system to improve a low quality image. The enhanced encoder system upsamples a low quality image and compares the upsampled image to a true high quality image to determine image inaccuracies that arise due to the upsampling process. The enhanced encoder system encodes the information describing the image inaccuracies using a trained encoder model as the residual bitstream. The image enhancement system upsamples the same low quality image to obtain a prediction of a high quality image that can include image inaccuracies. Given the residual bitstream, the image enhancement system decodes the residual bitstream using a trained decoder model and uses the additional image information to improve the predicted high quality image. The image enhancement system can provide an improved, high quality image for display.

Growing community of inventors

Mountain View, CA, United States of America

Oren Rippel