Parallel Synapses with Transmission Nonlinearities Enhance Neuronal Classification Capacity.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Cortical neurons utilize multiple synaptic contacts per postsynaptic neuron.
• Functional redundancy is avoided by distinct computational properties of parallel synapses.

Purpose of the Study:

• To model and evaluate the computational enhancement provided by nonlinear parallel synapses in cortical neurons.
• To assess the impact of synapse number and learnable parameters on neuronal classification capacity.

Main Methods:

• Modeling synaptic current as a sigmoidal transmission function with learnable parameters (amplitude, slope, threshold).
• Evaluating neuron classification capacity with nonlinear parallel synapses compared to the Perceptron.
• Applying the model to a feedforward neural network for MNIST image classification.

Main Results:

• Neurons with nonlinear parallel synapses show substantially improved classification capacity over the Perceptron.
• Classification accuracy increases superlinearly with the number of presynaptic axons.
• The model neuron can implement arbitrary monotonic aggregate transmission functions.
• Application to MNIST classification demonstrated increased test accuracy.

Conclusions:

• Multiple nonlinear synapses per input axon significantly enhance a neuron's computational power.
• This synaptic architecture offers a powerful mechanism for neural computation and learning.
• The model provides insights into efficient information processing in biological and artificial neural networks.