Presentation
UnrealNAS: do neural architecture search with no labels
DescriptionThis work proposes NoisyQuant, a noisy bias-enhanced post-training activation quantization method for the complicated activation distribution of vision transformer models. We theoretically prove the quantization error reduction brought by adding noisy bias to the input activation before quantization, and empirically show the effectiveness of NoisyQuant on both linear and nonlinear quantizers on vision transformer models.
We hope this work opens up a new direction to reduce PTQ quantization error through actively changing the distribution being quantized, and inspires better Noisy Bias distribution design or generation method based on quantizer and activation characteristics.
We hope this work opens up a new direction to reduce PTQ quantization error through actively changing the distribution being quantized, and inspires better Noisy Bias distribution design or generation method based on quantizer and activation characteristics.
Event Type
Work-in-Progress Poster
TimeWednesday, July 12th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
RISC-V
Security
