Bitwidth-Adaptive Quantization-Aware Neural Network Training: A Meta-Learning Approach

AbstractDeep neural network quantization with adaptive bitwidths has gained increasing attention due to the ease of model deployment on various platforms different resource budgets. In this paper, we propose a meta-learning approach achieve goal. Specifically, MEBQAT, simple yet effective way bitwidth-adaptive quantization-aware training (QAT) where is effectively combined QAT by redefining tasks incorporate bitwidths. After being deployed platform, MEBQAT allows (meta-)trained be quantized any candidate bitwidth minimal inference accuracy drop. Moreover, in few-shot learning scenario, can also adapt as well unseen target classes adding conventional optimization or metric-based meta-learning. We design variants support both (1) scenario and (2) new are jointly adapted. Our experiments show that merging into results remarkable performance improvement: 98.7% less storage cost compared bitwidth-dedicated 94.7% back propagation bitwidth-only adaptation scenarios, while improving classification up 63.6% vanilla bitwidth-class joint scenarios.