Reproduce “Non-linear motor control by local learning in spiking neural networks” on OpenResearch

About this paper

Aditya Gilra, Wulfram Gerstner

A spiking neural network trained with the Feedback-based Online Local Learning Of Weights (FOLLOW) scheme at EPFL learns the inverse dynamics of a non-linear two-link robotic arm, enabling it to control the arm to accurately follow desired trajectories. The approach utilizes a differential feedforward architecture and a local learning rule to address the credit assignment problem in hidden layers, achieving robust performance when combined with a linear state-error feedback mechanism.

Key insights

• The differential feedforward architecture is highly effective for approximating temporal derivatives, a crucial operation for learning non-linear inverse dynamics with spiking neural networks.
• The FOLLOW learning scheme successfully resolves the credit assignment problem for hidden layers in SNNs, enabling stable and online learning of complex, non-linear mappings using only local synaptic updates.
• Combining a learned non-linear inverse model (feedforward control) with a simple linear feedback mechanism (state-error correction) provides a robust and efficient adaptive control strategy for complex dynamical systems.

Summary by alphaXiv.