ECQ$$^{\text {x}}$$: Explainability-Driven Quantization for Low-Bit and Sparse DNNs

Abstract The remarkable success of deep neural networks (DNNs) in various applications is accompanied by a significant increase network parameters and arithmetic operations. Such increases memory computational demands make learning prohibitive for resource-constrained hardware platforms such as mobile devices. Recent efforts aim to reduce these overheads, while preserving model performance much possible, include parameter reduction techniques, quantization, lossless compression techniques. In this chapter, we develop describe novel quantization paradigm DNNs: Our method leverages concepts explainable AI (XAI) information theory: Instead assigning weight values based on their distances the clusters, assignment function additionally considers relevances obtained from Layer-wise Relevance Propagation (LRP) content clusters (entropy optimization). ultimate goal preserve most relevant weights highest content. Experimental results show that Entropy-Constrained XAI-adjusted Quantization (ECQ $$^{\text {x}}$$ x ) generates ultra low-precision (2–5 bit) simultaneously sparse maintaining or even improving performance. Due reduced precision high number zero-elements, rendered are highly compressible terms file size, up 103 $$\times $$ × compared full-precision unquantized DNN model. approach was evaluated different types models datasets (including Google Speech Commands, CIFAR-10 Pascal VOC) with previous work.