Spiking representation learning for associative memories.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking neural networks (SNNs) are biologically inspired computational models.
• Current SNNs face challenges in scaling and real-world data processing compared to deep learning.
• Effective learning of distributed representations is crucial for SNNs' perceptual and cognitive functions.

Purpose of the Study:

• Introduce a novel SNN architecture for unsupervised representation learning and associative memory.
• Address limitations in scaling and real-world applicability of current SNN models.
• Leverage biologically plausible plasticity mechanisms for enhanced SNN performance.

Main Methods:

• Developed a novel SNN utilizing Hebbian synaptic and activity-dependent structural plasticity.
• Modeled neuron units as Poisson spike generators with sparse firing rates.
• Designed a neocortical columnar-inspired architecture with feedforward and recurrent projections.

Main Results:

• The SNN successfully performed unsupervised representation learning.
• The model demonstrated associative memory functions, including pattern completion and prototype extraction.
• Evaluated performance on attractor-based memory properties like distortion resistance and perceptual rivalry.

Conclusions:

• The proposed SNN model offers a promising approach for unsupervised learning and associative memory.
• The architecture and plasticity rules provide a biologically plausible framework for advanced neural computation.
• This work advances the potential of SNNs for complex real-world applications.