An Oscillatory Neural Autoencoder Based on Frequency Modulation and Multiplexing | JoVE Visualize

Area of Science:

Computational Neuroscience
Artificial Neural Networks
Signal Processing

Background:

Oscillatory phenomena are fundamental to brain function, yet oscillator-based neural models are underexplored compared to rate-coded and spiking models.
Existing oscillator models are often limited to specific functions like rhythm generation or memory, lacking broader applicability to diverse neural dynamics.
Autoencoders, crucial for deep learning, have been developed using rate-coded and spiking neurons, but not yet with oscillatory principles.

Purpose of the Study:

To propose and develop a novel oscillatory neural network model capable of performing autoencoder functions.
To create a biologically inspired model that integrates rate-coded neurons and neural oscillators for enhanced computational capabilities.
To demonstrate the model's efficacy in processing and compressing time series data, including real-world electroencephalogram (EEG) signals.

Main Methods:

Developed a hybrid neural network combining rate-coded neurons and neural oscillators.
Input signals modulate encoder oscillator frequencies, which are then multiplexed by a Lateral Anti-Hebbian Network (LAHN).
Signal reconstruction is achieved through a de-multiplexing output layer using adaptive Hopf and Kuramoto oscillators, forming a neural phase-locked loop.

Main Results:

The proposed oscillatory autoencoder model successfully performs signal reconstruction.
The model demonstrates applicability to time series data, showing effective compression of synthetic and real-world EEG signals.
The Kuramoto-Hopf oscillator combination provides a novel mechanism for neural demodulation, analogous to a phase-locked loop.

Conclusions:

The developed oscillatory neural network serves as a functional autoencoder, expanding the repertoire of biologically inspired neural network models.
This hybrid model offers a promising approach for processing and compressing complex time series data, particularly in neuroscience applications like EEG analysis.
The integration of oscillatory dynamics within an autoencoder framework opens new avenues for understanding brain computation and developing advanced neural network architectures.