Quantization with Guaranteed Floating-Point Neural Network Classifications

Learn about a novel approach to neural network quantization that maintains classification consistency with original floating-point networks in this 15-minute conference presentation from OOPSLA 2025. Discover how researchers Anan Kabaha and Dana Drachsler Cohen from Technion address the critical challenge of computational cost reduction in neural networks while preserving classification accuracy. Explore the CoMPAQt algorithm, which uses mixed-integer linear programming (MILP) with custom linear relaxations to compute maximal classification confidence and detect inconsistencies at inference time. Understand two correction mechanisms: one that guarantees 100% consistency with floating-point networks using increasing bit precisions, and another that uses ensemble methods to mitigate classification inconsistencies. Examine experimental results on MNIST, ACAS-Xu, and tabular datasets showing 3.8x to 4.1x computational cost reductions while maintaining near-perfect classification consistency. Gain insights into the first formal guarantee approach for classification consistency in quantized neural networks, combining verification techniques with practical quantization methods for both fully connected and convolutional architectures.