CerebelluMorphic: Large-Scale Neuromorphic Model and Architecture for Supervised Motor Learning

Area of Science:

• Neuroscience
• Neuromorphic Engineering
• Computational Neuroscience

Background:

• The cerebellum is crucial for motor control and learning, utilizing supervised learning principles.
• Neuromorphic engineering aims to replicate biological neural systems for advanced computation.
• Existing models often lack the scale and biological fidelity of the cerebellum.

Purpose of the Study:

• To present a large-scale cerebellar network model for supervised learning.
• To develop a cerebellum-inspired neuromorphic architecture mirroring its anatomical structure.
• To enhance biomimicry and computational performance in brain-inspired systems.

Main Methods:

• Construction of a multinucleus cerebellar model with ~3.5 million neurons and 3411k granule cells.
• Mapping the anatomical structure onto a large-scale neuromorphic architecture.
• Implementation on a reconfigurable neuromorphic system to replicate cerebellar dynamics.

Main Results:

• The model achieves a 34x upscaling in neurons and a 284x increase in granule cells compared to prior work.
• Biologically plausible divergence/convergence ratios were achieved, improving biological mimicry.
• Real-time operation demonstrated with 4.70x higher throughput than previous studies.

Conclusions:

• The proposed model and architecture offer a robust platform for brain-inspired computing.
• It provides a theoretical basis and engineering perspective for cerebellar learning research.
• The work advances the exploration of neuromorphic systems for complex cognitive functions.