VisionHOPE: a self-modifying visual backbone from Min离 and CAS hits 85.2% on ImageNet
机器之心 · wechat · 2026-10-18
Min色 Technology and CAS's Institute of Automation propose VisionHOPE, the first general visual backbone built around a self-modifying learning system. Its SRNL module maintains five updatable memories (content, Key, Value, learning-rate, retention) so learning rules update within a single forward pass, with Delta gradient descent to avoid redundant writes and two provable stability constraints. A four-directional scan with 14-token chunking gives 2.86x inference speedup over per-token processing at only 0.1pp accuracy cost. VisionHOPE-S (52.9M params) reaches 85.2% Top-1 on ImageNet-1K — beating VMamba-B with under 60% of its parameters — and is best-in-class on COCO and ADE20K. Code, CUDA operators and weights are MIT-licensed; HF daily papers #2, 1000 GitHub stars. Memory still resets per image, so cross-frame continual learning remains future work.
More from Research
- Domingos Cites New Result Claiming RL Is 'Basically Useless' vs Supervised Learning — pmddomingos · 2026-10-23
- Wolfram Opens Rule 30 Prizes to AIs: $30,000 for Three Open Problems — cephaloform · 2026-10-23
- Stanford and Databricks use AI agents to reproduce medical findings and test analytical robustness — ChengleiSi · 2026-10-23
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23