VisionHOPE: Visual Backbones as Self-Modifying Learning Systems

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! Series: Weekly AI Papers · Recent research VisionHOPE explores how a vision model can change both its memory and its learning rules while processing an image. Its design connects five memories: content, keys, values, learning rate, and retention. The paper develops controls that keep these coupled updates stable under the conditions established by its analysis. Four directional scans adapt the method to the two-dimensional structure of images, with computation scaling linearly in the number of image tokens. The authors report competitive results on ImageNet classification, COCO object detection, and ADE20K semantic segmentation. This 2026 preprint connects associative memory and Nested Learning to the design of visual backbones. Source: VisionHOPE: Visual Backbones as Self-Modifying Learning Systems · Siran Peng, Tianshuo Zhang et al. Explore: AI & Computer Science · AI Papers Related episodes: V-JEPA (LeCun) · Machines That Think Like People (Lake et al.)