SpikePropamine: Differentiable Plasticity in Spiking Neural Networks

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Biological neural networks utilize adaptive synaptic efficacy for learning.
• Current Spiking Neural Networks (SNNs) often have static synapses, limiting post-training adaptation.
• This limitation hinders SNNs' ability to perform continuous learning and adapt to new information.

Purpose of the Study:

• To develop a framework for simultaneously learning fixed synaptic weights and the rules of synaptic plasticity in SNNs.
• To enable dynamic and continuous learning in SNNs through gradient-based optimization.
• To investigate the effectiveness of differentiable plasticity in SNNs for complex temporal tasks and robotic control.

Main Methods:

• Introduced a novel framework for end-to-end learning of synaptic plasticity rules in SNNs.
• Utilized gradient descent to optimize both fixed weights and plasticity parameters.
• Evaluated the framework on challenging temporal learning benchmarks and a robotic locomotion task.

Main Results:

• The proposed framework successfully learned parameters for various plasticity rules, including BCM and Oja's, and their neuromodulatory variants.
• SNNs with differentiable plasticity outperformed traditional SNNs on temporal learning tasks, even with high noise levels.
• The networks demonstrated robust performance in producing robotic locomotion, showing minimal degradation in novel conditions.

Conclusions:

• Differentiable synaptic plasticity offers a powerful mechanism for enhancing SNN learning capabilities.
• This framework allows SNNs to perform continuous adaptation and generalize to unseen situations.
• The approach holds promise for developing more adaptive and resilient artificial intelligence systems.