Demonstrating Advantages of Neuromorphic Computation: A Pilot Study.

Area of Science:

• Neuroscience and Computer Engineering
• Development of brain-inspired computing systems

Background:

• Neuromorphic devices aim to replicate brain functionality for enhanced computation, learning, and energy efficiency.
• Spiking neural networks are a key component in neuromorphic engineering.

Purpose of the Study:

• To demonstrate reward-modulated spike-timing-dependent plasticity (R-STDP) on a neuromorphic system.
• To implement a proof-of-concept learning agent for a simplified game using on-chip learning.

Main Methods:

• Utilized the BrainScaleS 2 neuromorphic system, a mixed-signal substrate with an embedded digital processor.
• Implemented R-STDP for a spiking network to learn a smooth pursuit task in a simulated Pong game.
• Leveraged analog emulation for accelerated neuronal dynamics and on-chip digital processor for learning and environment simulation.

Main Results:

• Achieved a 1000x acceleration in biological real-time with a 57 mW power budget.
• Demonstrated on-chip learning that mitigates fixed-pattern noise and utilizes temporal variability.
• Showcased synaptic weight adaptation to compensate for neuronal parameter variability.

Conclusions:

• Neuromorphic hardware can achieve superior speed and energy efficiency compared to software simulations.
• On-chip learning effectively compensates for hardware imperfections in analog neuromorphic substrates.
• This work validates the potential of neuromorphic systems for real-time, adaptive learning tasks.