Learning real-world stimuli in a neural network with spike-driven synaptic dynamics.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Hardware Implementation

Background:

• Existing artificial neural networks (ANNs) face limitations in real-time learning and energy efficiency.
• Biologically plausible models of synaptic plasticity offer potential for more robust and efficient learning systems.
• The need for hardware capable of semisupervised learning of complex stimuli is growing.

Purpose of the Study:

• To present a novel model of spike-driven synaptic plasticity inspired by experimental neuroscience.
• To develop an electronic hardware device capable of semisupervised learning and complex stimuli classification.
• To validate the model's performance against established machine learning benchmarks.

Main Methods:

• Utilized integrate-and-fire neurons with constant leak and a floor.
• Implemented bistable synapses modified by presynaptic spikes, influenced by postsynaptic depolarization and integrated action potentials.
• Employed a semisupervised training approach with an instructor signal, followed by unsupervised recall.

Main Results:

• The network demonstrated robust memory retention and resilience to spontaneous activity.
• A 2000-input neuron network successfully classified thousands of highly overlapping patterns (e.g., NIST characters).
• Achieved classification and generalization performance comparable to or better than ANNs.

Conclusions:

• The proposed synaptic plasticity model is suitable for building efficient, learning electronic hardware.
• The model's dynamics align with experimental observations of long-term synaptic modifications, including spike-timing-dependent plasticity.
• This biologically inspired approach offers a promising alternative for complex pattern classification tasks.