News ·
ExPLoRe accepted to ECCV 2026
The lab's latest self-supervised learning work routes per-patch training signal through Soft-MoE dispatch weights — closing out a three-paper arc from Cross-Scale MAE through MEDiC.
ExPLoRe: Expert Patch-Level Loss Routing for Multi-Objective Masked Image Modeling, by Konstantinos Georgiou, Maofeng Tang, and Dr. Qi, has been accepted to the European Conference on Computer Vision (ECCV 2026).
ExPLoRe resolves a tension the lab’s earlier MEDiC work exposed: when several self-supervised objectives share one vision transformer, different image patches want different training signals. The method repurposes the dispatch weights of a Soft mixture-of-experts as learned per-patch loss coefficients, reaching 80.6% linear-probe and 85.3% fine-tune accuracy on ImageNet-1K with a ViT-B/16.
Code is available in the lab’s GitHub repository, and the full research arc is described on the representation learning project page.