Competitive Learning in a Spiking Neural Network: Towards an Intelligent Pattern Classifier.

Area of Science:

• Neuroscience and Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking neuron networks (SNNs) are increasingly used in human-machine interfaces for signal processing.
• SNNs can be trained using biologically inspired algorithms, offering efficient learning.
• Encoding information via spiking frequency and spike timing presents a challenge for SNN learning algorithms.

Purpose of the Study:

• To investigate how single neurons and SNNs can be trained using mixed temporal-rate coding.
• To develop a universal SNN capable of learning from both temporal and rate-based input patterns.
• To evaluate the effectiveness of Hebbian learning and synaptic competition for SNN training.

Main Methods:

• Studied Hebbian learning rules (pair-based and triplet-based spike timing-dependent plasticity - STDP).
• Incorporated synaptic competition via a neuron-activity-dependent forgetting function.
• Developed unsupervised and supervised SNNs for electromyographical (EMG) pattern classification.

Main Results:

• Hebbian learning (STDP) is effective for temporal coding but not rate coding in SNNs.
• Synaptic competition is necessary for neural selectivity in rate coding.
• A hybrid approach using triplet-based STDP and synaptic competition successfully enabled rate coding.
• The proposed SNN achieved high accuracy (median 99.5%) in classifying EMG patterns using supervised learning.

Conclusions:

• A combination of STDP and synaptic competition provides a robust learning mechanism for SNNs handling mixed coding.
• SNNs can effectively classify complex biological signals like EMG patterns.
• The developed SNNs demonstrate a viable alternative to traditional machine learning models like multi-layer perceptrons for specific tasks.