Minimal Spiking Neuron for Solving Multilabel Classification Tasks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• The multispike tempotron (MST) is a powerful but computationally expensive single spiking neuron model for supervised classification.
• Its complexity limits its suitability for efficient simulation and neuromorphic hardware implementation.

Purpose of the Study:

• To investigate the possibility of simplifying the MST model while preserving its learning and information processing capabilities.
• To introduce and evaluate a new family of generalized neuron models (GNMs) as a simpler alternative.

Main Methods:

• Introduced a family of generalized neuron models (GNMs), a subset of the spike response model.
• Compared the learning performance of GNMs against the MST across a range of parameters.
• Analyzed the key components of the GNM responsible for its classification abilities.

Main Results:

• GNMs demonstrate comparable learning performance to the MST over a wide parameter range.
• The temporal autocorrelation of the membrane potential is identified as crucial for the GNM's spatiotemporal pattern classification.
• The GNM can be interpreted as a chemical system, linking neural computation with molecular information processing.

Conclusions:

• Simplified generalized neuron models (GNMs) can effectively perform complex classification tasks, matching the performance of the more complex multispike tempotron (MST).
• The GNM's ability to learn is significantly influenced by membrane potential's temporal autocorrelation, enabling spatiotemporal pattern recognition.
• Proposed alternative training methods like error trace learning and backpropagation for GNMs, enhancing their practical applicability.