Paper accepted at IJCAI 2026!

We are thrilled to share that our paper, "Energy-Regularized Spatial Masking: A Novel Approach to Enhancing Robustness and Interpretability in Vision Models," has been accepted to IJCAI International Joint Conferences on Artificial Intelligence Organization this year!

We discovered that efficient, interpretable visual representations can emerge naturally from a principle of 𝗲𝗻𝗲𝗿𝗴𝘆 𝗺𝗶𝗻𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻. Our approach, 𝗘𝗻𝗲𝗿𝗴𝘆-𝗥𝗲𝗴𝘂𝗹𝗮𝗿𝗶𝘇𝗲𝗱 𝗦𝗽𝗮𝘁𝗶𝗮𝗹 𝗠𝗮𝘀𝗸𝗶𝗻𝗴 (𝗘𝗥𝗦𝗠), reframes feature selection as a differentiable energy problem, assigning each visual token a cost from two competing forces: an intrinsic importance term and a spatial coherence penalty.

𝗞𝗲𝘆 𝗵𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀:
✅ 𝗥𝗼𝗯𝘂𝘀𝘁𝗻𝗲𝘀𝘀: Energy-guided patch removal actually improves accuracy after discarding the highest-energy patches, revealing ERSM as an implicit denoising mechanism.
✅ 𝗜𝗻𝘁𝗲𝗿𝗽𝗿𝗲𝘁𝗮𝗯𝗹𝗲: Produces contiguous, object-aligned spatial masks that isolate semantic regions without any segmentation supervision not post-hoc saliency, but direct projections of the model's decision bottleneck.
✅ 𝗠𝗼𝗱𝗲𝗹 𝗔𝗴𝗻𝗼𝘀𝘁𝗶𝗰: A lightweight, plug-and-play Energy-Mask Layer that drops into standard convolutional backbones without architectural changes.
✅ 𝗘𝗺𝗲𝗿𝗴𝗲𝗻𝘁 𝗦𝗽𝗮𝗿𝘀𝗶𝘁𝘆: No fixed pruning budget, the network autonomously discovers its own information-density equilibrium.