Cognitive Mapping and Memory Systems in Neural Networks

Neural networks have emerged as powerful tools for modelling the brain’s capacity to form internal representations of space and relational knowledge, often termed cognitive maps. In biological systems, the hippocampus and entorhinal cortex collaboratively encode spatial layouts, temporal sequences and abstract associations, supporting navigation, episodic memory and inference. Artificial models inspired by these circuits employ mechanisms such as predictive coding, self-attention and continuous attractor dynamics to learn latent representations that mirror spatial distances, environmental landmarks and relational structures. Such frameworks enable vector navigation—computing goal-directed trajectories—and flexible generalisation across contexts by balancing integration of related experiences with separation of distinct episodes. Advances in network architectures now allow cognitive maps to extend beyond vision to multimodal inputs, offering unified algorithms for mapping auditory, tactile or linguistic domains. These insights hold significant promise for robotics, autonomous vehicles and understanding impairments in memory disorders.

Research from Nature Portfolio

Recent studies have demonstrated that predictive coding networks equipped with self-attention modules can autonomously construct spatially coherent latent maps while solving next-image prediction tasks. During navigation in virtual environments, these models develop vectorised encodings that quantitatively reflect inter-landmark distances and support direct goal localisation using visual inputs alone. Parallel work in human neuroscience has revealed that the hippocampus simultaneously maintains spatial and transition-based predictive maps, with orbitofrontal feedback dynamically reshaping these representations to guide reward generalisation. Over the course of learning, spatial maps strengthen as predictive maps wane, illustrating how neural systems flexibly select and update the most relevant cognitive map for inference and decision-making.

Cognitive Mapping and Memory Systems in Neural Networks publication trend

The graph below shows the total number of articles in cognitive mapping and memory systems in neural networks across all publications each year (not limited to Nature Index journals).

Technical terms

Cognitive map: An internal representation of spatial or relational structure that supports navigation and inference.

Predictive coding: A neural algorithm in which models generate predictions of sensory input and adjust representations based on residual errors.

Grid cell: A neuron in the entorhinal cortex whose firing fields form a regular, periodic lattice of locations in an environment.

Continuous attractor network: A recurrent network model that maintains stable patterns of activity corresponding to continuous variables, such as position or orientation.

Vector navigation: A strategy for locomotion in which an agent computes and follows a direct displacement vector from its current location to a goal.

References

1. Automated construction of cognitive maps with visual predictive coding. Nature Machine Intelligence (2024).
2. Hippocampal spatio-predictive cognitive maps adaptively guide reward generalization. Nature Neuroscience (2023).
3. A Review of Brain-Inspired Cognition and Navigation Technology for Mobile Robots. Cyborg and Bionic Systems (2024).
4. Accurate Path Integration in Continuous Attractor Network Models of Grid Cells. PLOS Computational Biology (2009).
5. The Tolman-Eichenbaum Machine: Unifying Space and Relational Memory through Generalization in the Hippocampal Formation. Cell (2020).