Complexity Reduction of Neural Lossless Data Compression via Cascade Probability Modeling

This paper addresses the computational inefficiency of neural lossless data compression methods, which typically suffer from two fundamental limitations: slow encoding and decoding speeds and fixed computational costs regardless of the compressed data complexity. We propose a novel neural complexity-adaptive cascade approach that trains compact, data-specific models and transmits their parameters alongside the encoded bitstream avoiding the need for retraining during the decoding. The proposed cascade architecture includes several relatively simple subnetworks called levels, so that each level refines the predictions of the previous levels. The main idea here is higher memory order of the input source means more levels should be used and vice versa. In order to define the number of levels needed for the encoding a stopping criterion that terminates further processing when additional compression gains fail to justify the computational overhead is proposed. Experimental results demonstrate that while maintaining competitive compression performance with an average performance degradation of only $6.7 \%$ across diverse datasets, our approach achieves significant speed improvements: $\mathbf{1 2 - 1 3 0 \%}$ faster compression and 64.8$266.7 \%$ faster decompression across diverse data types. The most substantial speedup is observed on poorly compressible data, such as floating-point numbers, where the method demonstrates its highest efficiency gains.