Quantization with Guaranteed Floating-Point Neural Network Classifications
summary
CoMPAQt computes a maximal classification-confidence bound using a novel MILP encoding of quantization plus customized linear relaxations and constraints that tie the quantized and floating-point networks. The offline bound lets an inference-time checker flag inputs and trigger corrections: adaptive-precision refinement guarantees 100% consistency while cutting average multiply cost by 3.8x, and an ensemble reaches near-perfect consistency with ~4.1x speedups; evaluated on MNIST, ACAS‑Xu and tabular models.
AI-written summary · about
chapters (12)
- 0:00Quantization challenges
- 1:23Classification inconsistencies
- 3:07Quiff bound concept
- 5:54Computing bound with MIL
- 7:20System design two-stage
- 8:23Adaptive precision correction
- 9:06Evaluation setup metrics
- 10:33APR performance results
- 11:18Air traffic collision case
- 12:19Summary contributions
- 13:05Q&A correction
- 14:31Q&A scalability
related talks
34mBeyond the Continuum: The Importance of Quantization in Deep Learning
23mTowards Bit-Level Dominance Preserving Quantization of Neural Classifiers
40mUnlocking Local LLMs with Quantization
59mVerification and Machine Learning
08mVerifying Low-dimensional Input Neural Networks via Input Quantization
01h 14mDistributed and concurrent optimization for machine learning
19mQuant-LLM: Accelerating the Serving of Large Language Models via FP6-Centric...
26mLow-Precision Inference without Quality Loss...