VisionHOPE open-sourced: 52.9M-param vision backbone hits 85.2% ImageNet Top-1 with self-modifying learning
jiqizhixin · x · 2026-10-27
VisionHOPE, a vision backbone from the Institute of Automation (CAS) and Minglue Technology, is now open source with 1,000+ GitHub stars and #2 on the Hugging Face papers leaderboard.
- Results: 52.9M parameters, 85.2% Top-1 on ImageNet-1K — under 60% of VMamba-B's 89M params yet 1.3 points higher; VisionHOPE-B reaches 85.6%.
- Core idea: existing backbones adapt what they compute per image, but the learning rules — how information is learned, how much to update, what to keep — stay frozen at training time. VisionHOPE lets the model modify its own learning rules within a single forward pass, mimicking how human perception lets early clues reshape later looking.
- The diagram traces the lineage from CNN local aggregation to self-referential rule updates.
More from Research
- Music Lenses: ISMIR paper uses sparse autoencoders for steerable concept-level music retrieval — umpedronosapato · 2026-10-27
- Neuronpedia adds 7M+ neurons for Qwen3-32B, Gemma and Llama 3.3 70B to its interpretability encyclopedia — aryaman2020 · 2026-10-27
- Matthew Green hails stunning new ETH result pending Lean proof validity — matthew_d_green · 2026-10-27
- Heuristic for prediction accuracy: a bimodal probability histogram peaking at 0 and 1 — moultano · 2026-10-27
- AI agent autonomously discovers and implements high-order Maxwell solver scheme — burny_tech · 2026-10-27
- OpenAI claims its AI invented a new math technique, cracking a 35-year-old wave-focusing problem — 141_1337 · 2026-10-27