Liquid State Machine on SpiNNaker for Spatio-Temporal Classification Tasks.

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering
• Artificial Intelligence

Background:

• Liquid State Machines (LSMs), utilizing Spiking Neural Networks (SNNs), offer biological plausibility and efficient training for pattern recognition.
• Implementing LSMs for complex tasks like event-based vision on large-scale neuromorphic hardware remains challenging.

Purpose of the Study:

• To demonstrate the efficacy of offline-trained LSMs on the SpiNNaker neuromorphic processor for event-based visual classification.
• To evaluate the impact of transferring LSMs from deep learning frameworks to neuromorphic hardware and assess the effect of weight quantization.

Main Methods:

• Offline training of LSMs using an adaptation of back-propagation-through-time (BPTT) for SNNs.
• Deployment of trained LSMs onto the SpiNNaker neuromorphic processor for visual event classification.
• Performance evaluation on the N-MNIST dataset and analysis of weight quantization effects.

Main Results:

• Achieved state-of-the-art performance in classifying visual events on the N-MNIST dataset.
• Demonstrated that mapping LSMs from deep learning frameworks to SpiNNaker does not degrade classification accuracy.
• Confirmed that weight quantization significantly reduces memory footprint with minimal impact on performance.

Conclusions:

• Offline-trained LSMs are a viable and high-performing solution for event-based visual processing on neuromorphic hardware.
• The SpiNNaker platform effectively supports LSMs for complex tasks, bridging deep learning frameworks and neuromorphic implementation.
• Weight quantization presents a practical method for optimizing LSMs for memory-constrained neuromorphic systems.