SQS: Bayesian DNN Compression through Sparse Quantized Sub-distributions

Researchers propose SQS, a Bayesian method that unifies pruning and quantization for deep neural network compression by modeling weights with sparse quantized sub-distributions, achieving higher compression rates while preserving accuracy. The method targets deployment on resource-constrained devices, addressing limitations of existing approaches that apply pruning or quantization separately.

Compressing large-scale neural networks is essential for deploying models on resource-constrained devices. Most existing methods adopt weight pruning or low-bit quantization individually, often resulting in suboptimal compression rates to preserve acceptable performance drops. We introduce a unified