Heterogeneous recurrent spiking neural network for spatio-temporal classification.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) are brain-inspired AI models with potential for efficient computation.
• While supervised SNNs achieve high accuracy, unsupervised SNNs lag significantly in performance.
• Existing unsupervised SNNs lack the architectural and learning heterogeneity observed in biological systems.

Purpose of the Study:

• To develop a novel unsupervised learning model for spatio-temporal classification in video activity recognition.
• To introduce a heterogeneous recurrent spiking neural network (HRSNN) architecture and training methodology.
• To evaluate the performance of HRSNN on both RGB and event-based video datasets.

Main Methods:

• Proposed a Heterogeneous Recurrent Spiking Neural Network (HRSNN) with neurons exhibiting diverse firing/relaxation dynamics.
• Implemented a heterogeneous Spike-Time-Dependent Plasticity (STDP) learning rule with synapse-specific learning rates.
• Trained and evaluated the HRSNN model on KTH, UCF11, UCF101 (RGB), and DVS128 Gesture (event-based) datasets.

Main Results:

• Achieved high classification accuracies: 94.32% (KTH), 79.58% (UCF11), 77.53% (UCF101), and 96.54% (DVS Gesture).
• Demonstrated superior performance compared to homogeneous spiking neural networks.
• HRSNN matched supervised SNN performance with reduced computational cost (fewer neurons, sparse connections) and less training data.

Conclusions:

• The proposed HRSNN model significantly advances unsupervised learning in Spiking Neural Networks for video activity recognition.
• Heterogeneity in both network architecture and synaptic plasticity is crucial for enhancing SNN performance.
• HRSNN offers a computationally efficient and data-efficient alternative to supervised SNNs and traditional deep networks.