EFFICIENT INITIALIZATION OF QUANTIZED NEURAL NETWORKS

摘要

A system for training a quantized neural network comprising a processing unit adapted for: applying a post-training quantization method to a set of full-precision weight values of a trained neural network to compute a set of quantized weight values; initializing a set of model weight values of a neural network using the set of quantized weight values; and training the neural network using a quantized training method to produce a set of trained quantized weight values.