Faster and stronger Lossless Compression with Optimized Autoregressive framework
DescriptionThis paper presents a novel general-purpose lossless compression framework that can be faster and stronger. By analyzing the AR-based compression process, we point out that the major redundancy is the repeated process problem. An individual-mix structure is proposed to address this issue. Moreover, we introduced a simple matrix multiplication operation to replace previous hardware-unfriendly Gumble-softmax sampling. This paper assigns unique parameters for each position in the batch. This allows the probability estimator to capture the distribution difference between different sub-sequences. Experiment results show proposed framework can achieve 130% speed improvement with 5% compression ratio improvement across data domains.
Event Type
Research Manuscript
TimeWednesday, July 12th10:40am - 10:55am PDT
Location3003, 3rd Floor
AI/ML Application and Infrastructure