Paper accepted at LCN 2026
Description
Our paper “LIQOS-ViT: Learned Importance guided Quantization with Outlier awareness for Split inference ViTs” has been accepted at LCN 2026, the 51st IEEE Conference on Local Computer Networks, to be held in Coimbra, Portugal, October 7-9, 2026.
Vision Transformers (ViTs) are powerful but too heavy to run fully on resource-constrained edge and IoT devices. Split computing addresses this by dividing inference between the edge device and a remote server, but transmitting intermediate activations across the split point can introduce substantial communication overhead. In this work, we propose an adaptive mixed-precision quantization framework that estimates the importance of individual activation values, groups and normalizes them to limit the impact of outliers, and allocates bit-widths accordingly under a global compression budget. The resulting approach reduces transmission cost while preserving model accuracy, making split ViT inference more practical in edge-server environments.
The paper is available here.