Deep oscillatory neural network.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Conventional neural networks possess static internal states, limiting their ability to model dynamic biological processes.
• Brain-inspired computing offers potential for more sophisticated and interpretable AI models.

Purpose of the Study:

• To propose a novel neural network architecture, the Deep Oscillatory Neural Network (DONN), incorporating brain-like oscillatory dynamics.
• To explore the application of DONN and its convolutional variant (OCNN) in signal and image processing tasks.
• To investigate the emergent properties and interpretability of oscillatory neural networks.

Main Methods:

• Developed DONN, integrating neural Hopf oscillators with complex-valued neurons (sigmoid, ReLU).
• Implemented three input signal modes for oscillators: resonator, amplitude modulation, and frequency modulation.
• Utilized complex backpropagation for training and extended the architecture to Oscillatory Convolutional Neural Networks (OCNNs).

Main Results:

• DONN and OCNN achieved comparable or superior performance on benchmark signal and image processing tasks.
• Observed emergent phenomena like feature and temporal binding during image classification.
• Demonstrated Spike Timing Dependent Plasticity (STDP) kernel behavior with Hebbian learning.

Conclusions:

• Deep Oscillatory Neural Networks offer a brain-inspired approach to AI with enhanced learning capabilities.
• The explicit oscillatory dynamics contribute to improved interpretability of internal network representations.
• DONN and OCNN show promise for advanced signal and image processing applications.