CEG4N: Counter-Example Guided Neural Network Quantization Refinement

EasyChair Preprint 8617

Abstract

Neural networks are essential components of learning-based software systems. However, their high compute, memory, and power requirements make using them in low resources domains challenging. For this reason, neural networks are often compressed before deployment. Existing compression techniques tend to degrade the network accuracy. We propose Counter-Example Guided Neural Network Compression Refinement (CEG4N). This technique combines search-based quantization and equivalence verification: the former minimizes the computational requirements, while the latter guarantees that the network’s output does not change after compression. We evaluate CEG4N on a diverse set of benchmarks that include large and small networks. Our technique successfully compressed 80% of the networks in our evaluation while producing models with up to 72% better accuracy than state-of-the-art techniques.

Keyphrases: Equivalence Verification, Neural Network Compression, Neural Network Equivalence, Robust Compression

@booklet{EasyChair:8617,
  author    = {João Batista Pereira Matos Júnior and Iury Bessa and Edoardo Manino and Xidan Song and Lucas C. Cordeiro},
  title     = {CEG4N: Counter-Example Guided Neural Network Quantization Refinement},
  howpublished = {EasyChair Preprint 8617},
  year      = {EasyChair, 2022}}