A new paper published in Nature Machine Intelligence details an artificial intelligence system that employs cognitive maps to facilitate goal-directed imagination and planning. This approach, inspired by how the human brain navigates and solves problems, offers a path toward more energy-efficient AI. The researchers argue that this brain-inspired method can overcome limitations in current AI systems, which often require extensive training and significant energy consumption.
The developed system, referred to as the Generative Cognitive Map Learner (GCML), builds an internal representation of how actions affect the world. It then mentally simulates various paths to reach a desired goal. Unlike many contemporary AI systems that rely on massive neural networks or reinforcement learning, GCML learns continuously through biologically plausible local learning rules. This allows it to adapt to new goals by imagining future action sequences rather than needing to memorize solutions.
Cognitive maps, broadly understood as data structures that organize learned experiences by encoding relational information between states and actions, are central to this research. While often studied in the context of spatial navigation in animals, the concept extends to abstract concept spaces. The brain uses these maps for planning and problem-solving, including for novel challenges beyond prior experience. Experimental data suggest that replay from spatial and non-spatial cognitive maps provides a neural basis for goal-directed imagination and compositional computation.
The researchers highlight that current AI often requires retraining for each new objective, a contrast to the brain's ability to adjust to changing circumstances. The GCML model addresses this by using stochastic neural computation, cognitive maps, and compositional coding. Its design is suitable for implementation in energy-efficient neuromorphic hardware, requiring only self-supervised local synaptic plasticity for on-chip learning. This contrasts with deep neural networks and large language models, suggesting that core features of intelligence, like generating solutions to unprecedented problems, can be achieved with less computational overhead.
The paper demonstrates the model's functional abilities through 2D spatial navigation tasks. Using a model based on the grid cell system, which employs path integration to update spatial location, the resulting cognitive map generates imagined trajectories towards any given spatial goal. These imagined paths resemble replay trajectories observed in rodent brains. The approach also involves controlled randomness in these imagined trajectories, enabling the generation of multiple candidate solutions that remain biased toward the desired outcome. This allows the system to discover alternative routes or strategies while still converging on a goal.
The researchers suggest that combining principles of cognitive maps, stochastic sampling, and compositional representations could enable energy-efficient AI systems capable of planning, problem-solving, and adapting to novel situations without the high computational costs associated with current deep learning methods. This brain-inspired approach aims to replicate core features of intelligence found in biological systems, such as continuous learning and rapid adaptation to new contingencies.
The work aligns with broader research trends in AI that seek to emulate biological intelligence for greater efficiency and capability. For example, Google Research has explored neural architectures for navigation that learn to map and plan. Other research has investigated how AI can learn from human imagination and memory formation processes. The development of brain-inspired machine learning aims to create systems that are more energy-efficient, adaptive, and capable of complex cognitive functions akin to biological intelligence.