Learning Precise Spike Train-to-Spike Train Transformations in Multilayer Feedforward Neuronal Networks.

Area of Science:

• Computational Neuroscience
• Artificial Neural Networks

Background:

• Spiking neural networks (SNNs) offer bio-realistic computation but learning precise spike timing remains challenging.
• Existing methods often rely on spike rates or probabilistic models, limiting temporal precision.

Purpose of the Study:

• To develop a synaptic weight update rule for learning temporally precise spike train-to-spike train transformations in feedforward SNNs.
• To generalize error backpropagation to deterministic SNNs using only spike timing information.

Main Methods:

• Proposed an error functional comparing output spike trains based on their impact on a virtual postsynaptic neuron.
• Introduced virtual weight assignment to spikes for perturbation analysis of spike times and weights.
• Utilized gradient descent with derived gradients for synaptic weight updates.

Main Results:

• Demonstrated the efficacy of the proposed learning framework through simulation experiments.
• The method successfully learns temporally precise spike train transformations.
• Observed asymmetries between excitatory and inhibitory synapses.

Conclusions:

• The derived synaptic weight update rule effectively facilitates learning precise spike train transformations in SNNs.
• The novel error functional and perturbation analysis provide a robust framework for SNN learning.
• Further investigation into synaptic property differences in SNNs is warranted.