Unsupervised post-training learning in spiking neural networks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• The human brain utilizes diverse learning strategies, but Spiking Neural Networks (SNNs) typically employ only one, like spike-timing-dependent plasticity (STDP).
• Conventional neural networks are trained and then fixed, lacking adaptability to new information post-training.

Purpose of the Study:

• To investigate if short-term plasticity (STP) can enable SNNs to learn post-training without altering synaptic weights.
• To enhance the biological plausibility and computational power of SNNs by integrating multiple learning rules.

Main Methods:

• Combined triplet STDP for long-term learning with STP for short-term, post-training learning.
• Developed two unsupervised learning pipelines for image classification, incorporating a dynamic synapse model into trained SNNs.

Main Results:

• The proposed method achieved higher classification accuracy compared to traditional training approaches.
• The SNNs demonstrated a faster convergence rate, indicating improved learning efficiency.
• Successfully demonstrated the feasibility of post-training learning in SNNs through the integration of STP.

Conclusions:

• Short-term plasticity (STP) can be effectively integrated into Spiking Neural Networks (SNNs) to enable learning after initial training.
• This approach enhances biological plausibility and computational performance, offering a new paradigm for neural network development.
• Future research should explore the application of this post-training learning concept to diverse challenges and datasets.