Techniques for general-purpose lossless data compression using a recurrent neural network
AssigneeDropbox, Inc.
InventorsDaniel R. Horn, Jongmin Baek, Anatoly Yakovenko
Life of this patent
Estimating the term…
Abstract
Techniques for general-purpose lossless data compression using a neural network including compressing an original content item to a baseline lossless compressed data format. The baseline lossless compressed data format is binarized to a binarized format. The binarized format is arithmetically coded based on probability estimates from a neural network probability estimator. The neural network probability estimator generates the probability estimates for current symbols of the binarized format to be arithmetically coded based on symbols of the binarized format that have already been arithmetically coded.
Claims20
A method performed by a computing system comprising one or more processors and memory, the method comprising:
- compressing an original content item to a baseline lossless compressed data format;
- binarizing the baseline lossless compressed data format to a binarized format;
- arithmetically coding the binarized format based on probability estimates from a recurrent neural network probability estimator;
- and
- wherein the recurrent neural network probability estimator generates probability estimates for current symbols of the binarized format to be arithmetically coded based on symbols of the binarized format that have already been arithmetically coded during the arithmetically coding the binarized format.
One or more non-transitory computer-readable media storing one or more programs for executing by one or more processors, the one or more programs comprising instructions configured for:
- compressing an original content item to a baseline lossless compressed data format;
- binarizing the baseline lossless compressed data format to a binarized format;
- arithmetically coding the binarized format based on probability estimates from a neural network probability estimator;
- and
- wherein the neural network probability estimator generates probability estimates for current symbols of the binarized format to be arithmetically coded
- based on symbols of the binarized format that have already been arithmetically coded during the arithmetically coding the binarized format.
A computing system, comprising:
- one or more processors;
- storage media;
- one or more programs stored in the storage media for execution by the one or more processors, the one or more processors comprising instructions for:
- storing a plurality of statistical bins in-memory;
- compressing an original content item to a baseline lossless compressed data format;
- binarizing the baseline lossless compressed data format to a binarized format;
- arithmetically coding symbols of the binarized format based on probability estimates from a neural network probability estimator;
- and
- wherein the neural network probability estimator generates probability estimates for each symbol of the symbols of the binarized format based on:
- a prediction of the each symbol generated by a recurrent neural network, and statistics maintained in a selected bin of the plurality of statistical bins selected for the each symbol.
Dependencies are inferred from the claim text.