Synthesis of neural networks for spatio-temporal spike pattern recognition and processing

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Neural engineering

Background:

• Large-scale neural computational platforms require algorithms for synthesizing neural structures to perform cognitive tasks.
• Existing methods like the Neural Engineering Framework (NEF) are effective for spike rate representations but require numerous neurons.
• There is a need for efficient neural network synthesis methods that utilize time-encoded neural signals and sparse neuron populations.

Purpose of the Study:

• To present a novel neural network synthesis method for time-encoded neural signals.
• To enable the creation of sparsely connected neural networks that efficiently perform cognitive tasks.
• To allow arbitrary specification of neuronal characteristics and optimize synaptic connectivity.

Main Methods:

• A synthesis method generating synaptic connectivity for neurons processing time-encoded signals.
• Incorporation of user-defined neuronal characteristics (axonal/dendritic delays, synaptic transfer functions).
• Optimization of dendritic weights to achieve the desired input-output relationship, supporting batch or online learning.

Main Results:

• Demonstration of a method that makes very sparse use of neurons.
• An extremely fast optimization process for computed dendritic weights.
• Successful generation of a neural network for speech recognition using spike-time encoding.

Conclusions:

• The developed method offers an efficient alternative to existing neural network synthesis techniques.
• It enables the creation of biologically plausible and computationally efficient neural networks.
• The method holds promise for advancing artificial intelligence and neural engineering applications.