RISE Research Radar

Computer Science Open House 2022-2025

2024

Inhibition: Efficient Attention for Transformers

Rickard Brännvall, Andrei Stoian

Summary

The Inhibitor: ReLU and Addition-Based Attention for Computationally Efficient Quantized Transformers. Quantized Transformers without dot-product and Softmax. Replace with ReLU activation. Reduced precision for efficiency.

Themes

machine-learningllm

Keywords

efficient transformers, quantization, ReLU attention, Inhibitor, homomorphic encryption

Poster

Inhibition: Efficient Attention for Transformers poster

Click image to open full size