Self-organizing networks learn to predict and navigate without pre-wired connections

Researchers Facundo Emina and Emilio Kropff present a theoretical framework in which continuous attractor neural networks (CANNs) self-organize to exhibit two key cognitive abilities: prospective coding (anticipating future states) and path integration (tracking position from motion). Rather than relying on pre-wired recurrent connectivity, the network learns through Hebbian plasticity, firing-rate adaptation, and global inhibition. The authors show that translationally invariant inputs naturally yield stable Gaussian-shaped feedforward weights. Crucially, anticipatory dynamics arise spontaneously in this feedforward architecture—the activity bump shifts forward without any recurrent excitatory collaterals. This predictive shift can be linearly amplified across multiple layers, mirroring anticipatory activity observed in the superficial entorhinal cortex.

When recurrent interactions are introduced, the network learns to self-sustain a moving bump of activity, enabling path integration. By modulating the network with a time-varying baseline current that encodes speed, its intrinsic velocity adjusts to function as a precise unidirectional path integrator. The work suggests that prospective coding and path integration are not manually engineered features but emerge naturally as equilibrium solutions of a self-organizing competitive network. This has significant implications for designing more brain-like AI systems and understanding neural computation in biological brains.