Towards Bit-Level Dominance Preserving Quantization of Neural Classifiers
summary
Method computes the minimum bit precision PR that guarantees the original predicted class remains unchanged for all inputs, giving a static, per-layer post-training quantization that is independent of IEEE formats. It leverages Rump’s dot-product error bound to turn dominance preservation into a nonlinear constraint and solves it with a trust-region interior-point optimizer. On experiments with dense layers (MNIST, CIFAR-like sets) it reduces 23-bit mantissas to PR ≈ 9–15 bits, saving ~35–60% memory while provably preserving class dominance. Limitations: only single dense layers so far, no activation or multi-layer composition handled yet.
AI-written summary · about
chapters (11)
- 0:00Research motivation
- 2:16Quantization techniques
- 5:30Dominance preservation
- 9:09Mathematical formulation
- 12:23Ramp’s error bound
- 14:11Nonlinear optimization
- 16:21Experimental setup
- 17:48Results on memory
- 18:53Scalability analysis
- 20:44Limitations and future
- 22:31Take-home messages
related talks
15mQuantization with Guaranteed Floating-Point Neural Network Classifications
34mBeyond the Continuum: The Importance of Quantization in Deep Learning
40mUnlocking Local LLMs with Quantization
16mDecDEC: A Systems Approach to Advancing Low-Bit LLM Quantization
48mPractical Quantization in Keras
01h 14mDistributed and concurrent optimization for machine learning
14mOcto: INT8 Training with Loss-aware Compensation and Backward Quantization for Tiny
08mVerifying Low-dimensional Input Neural Networks via Input Quantization