Spiking neurons with short-term synaptic plasticity form superior generative networks

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) are explored for cortical computation and machine learning, but their superiority over non-spiking networks is debated.
• Classical machine learning algorithms face challenges with high-dimensional data, often requiring computationally intensive tempering techniques for sampling.

Purpose of the Study:

• To investigate the computational advantages of short-term synaptic plasticity (STSP) in SNNs.
• To demonstrate how STSP enables SNNs to overcome sampling problems in deep attractors.
• To compare the performance of SNNs with STSP against classical tempering techniques.

Main Methods:

• Developing SNNs incorporating local short-term synaptic plasticity.
• Evaluating network performance on high-dimensional, diverse, and imbalanced datasets.
• Analyzing the impact of biologically inspired, spike-triggered synaptic dynamics.

Main Results:

• SNNs with STSP can achieve results comparable to classical tempering techniques for sampling problems.
• These SNNs demonstrate superior performance over tempering-based approaches when dealing with imbalanced training data.
• Local STSP provides distinct computational advantages for processing complex sensory data.

Conclusions:

• Short-term synaptic plasticity confers unique computational benefits to spiking neural networks.
• Biologically inspired synaptic dynamics offer efficient solutions for complex machine learning tasks, particularly with imbalanced data.
• STSP in SNNs presents a promising avenue for advanced AI and neuroscience modeling.