CEREBELLAR FEEDFORWARD SUPPORT IMPROVES LEARNING EFFICIENCY AND CAPACITY IN RECURRENT CORTICAL NETWORKS

Complex cognition necessitates cortical circuits to balance adaptive flexibility and preservation of existing representations. A proposed resolution to this stability-plasticity dilemma is in distributed learning systems supporting the cortex. Increasing evidence highlights the importance of cortico-cerebellar interactions in non-motor cognitive functions. However, the computational principles governing cerebellar support for cortical cognition remain inadequately defined. This study aimed to clarify how the cerebellum aids cortical learning in complex cognitive functions through a biologically-inspired computational framework. To that end, we built a cerebellar feedforward computational model, with characteristic granule cell-like expansion layer, providing adaptive biases to a cortical recurrent neural network. We trained the model on complex memory tasks, employing curriculum learning to progressively increase task difficulty. Learning was organised into alternating phases: the cortical network was first trained to an accuracy threshold on the simplest variant, then frozen, allowing the cerebellum to adapt for the next difficulty level. Once performance met the threshold, cortical learning consolidated the solution before repeating the process. This structure enabled the cerebellar module to facilitate adaptive behaviour in response to changing task demands while minimising destabilising plasticity in the cortex, activating only when an optimal solution was identified. This model demonstrated more efficient learning and achieved higher levels of complexity than control models lacking cerebellar input, which were matched for size, learning capacity, or trained without structured alternation. These findings suggest that cerebellar learning can scaffold cortical learning through exploration while ensuring stable cortical consolidation, thereby providing a framework to better understand cortico-cerebellar interactions for cognition.