Co-learning synaptic delays, weights and adaptation in spiking neural networks

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Neuromorphic Engineering

Background:

• Spiking neural networks (SNNs) offer power-efficient, temporal processing advantages over artificial neural networks (ANNs) for neuromorphic hardware.
• Enhancing SNNs with biologically inspired features can improve their performance on complex tasks.

Purpose of the Study:

• To investigate the impact of co-learning synaptic weights with neuronal adaptation and synaptic propagation delays in SNNs.
• To evaluate the performance of these enhanced SNNs against baseline SNNs and ANNs on speech recognition tasks.

Main Methods:

• Co-learning synaptic weights with neuronal adaptation parameters to enable neurons to learn from their past activity.
• Co-learning synaptic weights with synaptic propagation delays to correlate temporally distant spike trains.
• Utilizing a simple 2-hidden layer feed-forward network architecture.

Main Results:

• Both neuronal adaptation and synaptic delays individually improved SNN performance over baseline.
• The combination of both features achieved state-of-the-art results on investigated speech recognition datasets.
• Enhanced SNNs outperformed benchmark ANNs, including GRUs, on neuromorphic and large-scale speech datasets, even with fewer parameters.

Conclusions:

• Co-learning biologically inspired features significantly enhances SNN capabilities for temporal processing.
• These brain-inspired improvements allow SNNs to surpass equivalent ANNs in tasks with rich temporal dynamics.
• The study demonstrates a promising direction for developing more efficient and powerful neuromorphic systems.