Nature Machine Intelligence paper shows backprop-free goal-directed planning
bravo_abad · x · 2026-07-22
This paper introduces GCML, a generative cognitive map learner that plans without backpropagation or backprop-through-time.
Core idea
- Learn a forward model plus a linear inverse model with local, Hebbian-style plasticity.
- Add Gaussian noise to action selection so goal-directed planning becomes stochastic sampling rather than expensive search.
- Given a target state, repeated noisy sampling produces diverse trajectories that can self-correct toward the goal.
Results
- 2D navigation: reproduces rodent hippocampal replay, including rerouting around novel barriers and exploring unseen space.
- Random graphs: acts as an online approximate k-shortest-path solver.
- NP-hard tiling task: trains on 5-block silhouettes and generalizes to 8-block shapes with near-perfect success.
Why it matters
- Compared with K, mA, BELA, RL, and MPC, GCML visits fewer nodes and avoids replanning overhead when the goal changes.
- Because updates are local, it may map well to in-memory or neuromorphic hardware.
- The authors suggest this could be useful for edge planning tasks such as materials screening, drug routing, or energy control.
Related event: Brain-Inspired GCML Achieves Planning Without Backpropagation(2 posts)→
More from Research
- STRIDE is being framed as a serious path toward neural reasoning with symbolic guarantees — williamtp · 2026-07-22
- Poolside’s Laguna S 2.1 118B release includes full eval trajectories, drawing praise — _lewtun · 2026-07-22
- Developer builds The Daily Diff, a curated AI and engineering news site — arpit_bhayani · 2026-07-22
- Agent memory needs ontology, write paths, and graph-based retrieval — Al_Grigor · 2026-07-22
- AI researchers say industry labs get millions while academia gets thousands — sokrypton · 2026-07-22
- VibeMathed tracks math problems AI models have helped prove or disprove — Abject_Response2855 · 2026-07-22