SpiLinC: Spiking Liquid-Ensemble Computing for Unsupervised Speech and Image Recognition.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Neuromorphic Engineering

Background:

• Liquid State Machines (LSMs) traditionally use fixed input-synaptic connections, limiting their effectiveness for complex pattern recognition.
• Supervised learning is typically required for readout layers in LSMs, hindering unsupervised applications.
• Existing LSMs are primarily demonstrated for speech recognition, with challenges in image pattern recognition.

Purpose of the Study:

• To propose a Spiking Neural Network (SNN) with plastic synapses for unsupervised speech and image recognition.
• To introduce an enhanced architecture, SpiLinC, addressing scalability and efficiency challenges in Liquid-SNNs.
• To enable direct inference of input class from neuronal spiking activity without supervised readout layers.

Main Methods:

• Developed a Liquid-Spiking Neural Network (Liquid-SNN) using Spike Timing Dependent Plasticity (STDP) for adaptive synaptic strengths.
• Implemented an unsupervised learning methodology for direct class inference from liquid neuronal activity.
• Proposed SpiLinC, an ensemble of parallel liquids employing a 'divide and learn' strategy for feature extraction.

Main Results:

• Liquid-SNN demonstrated efficient recognition of both speech and image patterns by learning temporal information.
• SpiLinC achieved competitive classification accuracy with improved sparsity and faster training convergence compared to Liquid-SNN.
• Both models were validated on the TI46 speech corpus (digits) and MNIST handwritten digits dataset.

Conclusions:

• The proposed Liquid-SNN and SpiLinC architectures offer effective unsupervised learning for complex pattern recognition tasks.
• SpiLinC enhances scalability and training efficiency, making it suitable for energy-efficient neuromorphic hardware.
• These models advance the application of SNNs beyond traditional LSMs for diverse recognition challenges.