Atlas neuron hits 96.7% after 20 tasks with no replay, labels or backprop
CatAstro_Piyush · x · 2026-10-01
A new post introduces the "Atlas neuron," tackling continual learning: can a model learn tasks sequentially without forgetting, with no replay, no task labels, and no backpropagation? On Permuted MNIST it retains 96.7% accuracy after 20 tasks, versus 45% for a sequentially trained MLP. Full details in the original thread.
More from Research
- HCOMP 2026 final session opens with FIBB paper on false information beliefs — windx0303 · 2026-10-01
- Meta-reasoning harness hits 71.5% on ProgramBench, beating Codex by 13.5 points — anirudhg9119 · 2026-10-01
- Rethinking inference-time scaling: what to compute, not just how much — anirudhg9119 · 2026-10-01
- Meta researchers propose Agentic Meta-Reasoning: letting models build their own reasoning graphs — anirudhg9119 · 2026-10-01
- FINGR: a dexterous hand solves a 2×2 Rubik's Cube with continuous finger tricks — HaozhiQ · 2026-10-01
- New lower bounds for kissing numbers: τ₁₉≥12268 and τ₂₁≥30761, open-sourced — felpix_ · 2026-10-01