Deep learning models typically use 32-bit floating-point numbers (FP32) for computation, incurring high computational cost and memory footprint. This makes them difficult to run efficiently on low-power hardware that only has low-precision integer arithmetic units.
This paper proposes a complete Quantization-Aware Training (QAT) scheme whose core goal is to enable neural network inference using only integer arithmetic, thereby significantly improving deployment efficiency on resource-constrained environments such as mobile devices and embedded systems.
It designs a quantization method that allows all weights and activations during the forward inference pass to be represented and computed as 8-bit integers (INT8), leveraging the efficiency of integer arithmetic while preserving the model's original accuracy as much as possible.