Neural sampling from cognitive maps enables goal-directed imagination and planning

2026-07-21 · Source: Nature Machine Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

A neural network model called Generative Cognitive Map Learner (GCML) integrates cognitive maps, stochastic computing, and compositional coding to enable goal-directed imagination and planning. This approach aims to achieve brain-like intelligence with high energy efficiency, contrasting with the high energy cost of current AI systems like deep neural networks and large language models. The GCML model utilizes self-supervised local synaptic plasticity for on-chip learning, making it suitable for neuromorphic hardware and edge devices. The model was validated across three distinct domains: reproducing rodent brain data on imagined spatial trajectories (including rerouting around novel barriers), solving problems in abstract concept spaces (such as finding k-shortest paths in random graphs), and tackling NP-hard compositional tasks like decomposing 2D silhouettes from building blocks. It demonstrates strong generalization to never-encountered states and offers diverse heuristic solutions with low computational effort and replanning latency compared to traditional Reinforcement Learning (RL) and Model Predictive Control (MPC) methods.

Key takeaway

For AI Scientists and Machine Learning Engineers developing energy-efficient solutions, you should explore neural sampling from cognitive maps, as demonstrated by the GCML model. This approach offers a path to flexible, goal-directed planning and problem-solving on neuromorphic hardware and edge devices, significantly reducing computational overhead compared to deep neural networks. Consider integrating local synaptic plasticity and stochasticity to generate diverse, adaptive solutions for novel tasks.

Key insights

A neural model integrates cognitive maps, stochasticity, and compositional computing for energy-efficient, flexible, goal-directed imagination and planning.

Principles

• Cognitive maps enable flexible planning and problem-solving.
• Stochasticity allows diverse, goal-directed solution sampling.
• Local synaptic plasticity supports online, on-chip learning.

Method

The Generative Cognitive Map Learner (GCML) integrates cognitive maps, stochastic computing, and compositional coding. It learns embeddings (Q, V) and an inverse model (W) via self-supervised local synaptic plasticity, then uses noise-infused action selection for goal-directed trajectory generation.

In practice

• Implement brain-inspired models on neuromorphic hardware.
• Develop energy-efficient AI for edge devices.
• Generate diverse solutions for complex planning tasks.

Topics

• Neural Sampling
• Cognitive Maps
• Neuromorphic Hardware
• Energy-Efficient AI
• Goal-Directed Planning
• Compositional Computing

Code references

• LH-cbicr/GCML

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Creating Intelligence: A Computational Foundation for AGI — A computational theory of mind grounds AGI in discrete set theory and topological plasticity for efficient, biologically-inspired cognition.
• Structured Representation Learning with Locally Linear Embeddings and Adaptive Feature Fusion — A neuroscientifically inspired RL framework disentangles dynamics and reward features using LLEs and adaptive fusion for improved learning.
• Artificial neurons successfully communicate with living brain cells — Printed artificial neurons can directly communicate with living brain cells, enabling energy-efficient brain-inspired computing.
• Hybrid ANN-SNN Pipeline with Local Plasticity — Hybrid ANN-SNN pipelines achieve high accuracy with local, biologically plausible learning by utilizing pretrained ANN encoders.
• Bridging the gap between neuromorphic ionic computing and more efficient AI — The brain's efficient neuromorphic computing inspires AI for better efficiency.
• Benchmarking the Energy Cost of Assurance in Neuromorphic Edge Robotics — Neuromorphic defenses can enhance AI robustness while simultaneously reducing energy consumption due to induced sparsity.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Nature Machine Intelligence.