A new recursive least squares-based learning algorithm for spiking neurons.

Area of Science:

• Computational Neuroscience
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) excel at processing spatio-temporal information.
• Developing effective supervised learning algorithms for SNNs remains a significant challenge due to complex temporal coding.

Purpose of the Study:

• To propose a novel supervised learning algorithm, the Recursive Least Squares-Based Learning Rule (RLSBLR), for SNNs.
• To enhance the generation of desired spatio-temporal spike trains using SNNs.

Main Methods:

• The RLSBLR updates weights based on a cost function comparing membrane potential to firing threshold.
• Weight modification considers both current and past error functions, incorporating a modified synaptic delay learning mechanism.
• Performance was evaluated across various experimental settings including spiking lengths, input numbers, firing rates, noise levels, and learning parameters.

Main Results:

• The RLSBLR achieved higher learning accuracy and efficiency compared to Perceptron-Based Spiking Neuron Learning Rule (PBSNLR) and Remote Supervised Method (ReSuMe).
• The algorithm demonstrated enhanced robustness against different types of noise.
• Application to the TIDIGITS database confirmed the RLSBLR's practical applicability and performance.

Conclusions:

• The proposed RLSBLR offers a significant advancement in supervised learning for SNNs.
• RLSBLR provides a more accurate, efficient, and robust method for spatio-temporal data processing with SNNs.
• The algorithm shows promise for real-world applications in areas like speech recognition.