AI accuracy

Output-Head Quantization via Softmax Reparameterization

A simple trick may improve output-head quantization by reducing final-layer errors; method, model, benchmarks,…