StaticBAC: Fast and Memory-Efficient Entropy Coding for Neural Network Compression

Abstract

The rapid growth in neural network model size has increased the need for efficient compression techniques, particularly for edge and resource-constrained environments. While quantization reduces model size, entropy coding can further compress already quantized parameters. State-of-the-art methods such as DeepCABAC achieve high compression efficiency through adaptive probability modeling but introduce significant computational and memory overhead in both the offline encoding and runtime decoding phases. In this work, we propose Static Binary Arithmetic Coding (StaticBAC), a non-adaptive entropy coding scheme that replaces adaptive context modeling with fixed probability models derived from offline analysis. This simplification reduces memory usage and control complexity, making StaticBAC suitable for hardware implementation in edge devices. Experimental results across five models show that StaticBAC achieves 88% of DeepCABAC compression, while providing up to $3 \times$ faster decoding and substantial memory savings, reducing memory usage by approximately 95% during encoding and 36% during decoding. These results demonstrate that static probability models offer a favorable trade-off between compression efficiency and implementation cost, enabling efficient compression for edgeAI. StaticBAC's code is available open source on GitHub: https://github.com/Jiovana/StaticBAC