Cerebellum-inspired RNN learns faster with modular architecture

A team led by Alexandra Voce, Emmanouil Giannakakis, and Claudia Clopath has developed a novel neural network architecture inspired by the brain's cortico-cerebellar circuits. Their CB-RNN (cortico-cerebellar RNN) augments a standard recurrent network with a feedforward module modeled after the cerebellum. In tests across varied temporal tasks, the CB-RNN learned significantly faster and reached higher peak performance compared to parameter-matched fully recurrent baselines. A key finding: after just minimal training of the recurrent core, freezing it and allowing only the cerebellar module to continue learning preserved the architecture's superior efficiency. This suggests the recurrent part functions largely as a fixed reservoir, with the cerebellar module driving rapid adaptation.

The work offers a concrete example of how heterogeneous modularity can serve as an inductive bias for temporal processing, a major challenge in AI. By mimicking the brain's separation of recurrent computation (cortex) and fast feedforward adjustment (cerebellum), the architecture achieves both speed and stability. For practitioners, this could lead to more sample-efficient sequence models and new design principles for neuromorphic hardware. The paper also reinforces the value of neuroscience-inspired AI, showing that biological constraints can yield practical algorithmic benefits beyond mere scaling.