CONE: Convex-Optimized-Synaptic Efficacies for Temporally Precise Spike Mapping.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) excel at time-dependent pattern recognition by utilizing precise spike timing.
• Supervised learning for SNNs is challenging due to complex, non-convex optimization landscapes in existing heuristic methods.
• Current SNN training methods may not guarantee convergence to optimal solutions.

Purpose of the Study:

• To develop a novel supervised learning technique for determining the weights of spiking neurons.
• To formulate the SNN weight optimization problem within a convex optimization framework.
• To analyze the impact of weight distribution and membrane trajectory on SNN robustness and memory capacity.

Main Methods:

• Formulation of spiking neuron weight determination as a convex optimization problem.
• Introduction of techniques to control weight distribution and membrane potential trajectories.
• Analysis of solution existence, robustness against noise, and memory capacity using synthetic data.
• Application to real-world gait-event detection using experimental data.

Main Results:

• A novel convex optimization framework for training spiking neurons was successfully developed.
• The method allows for control over neuron dynamics, enhancing robustness to noise.
• The technique was validated through synthetic examples and practical application in gait-event detection.

Conclusions:

• The proposed convex optimization approach offers a principled and effective method for training spiking neurons.
• This framework addresses limitations of heuristic methods, ensuring convergence and enabling analysis of SNN properties.
• The technique demonstrates practical utility in real-world applications like event detection from experimental data.