Digital Organoid: A Biologically Inspired Neural Network for Image Classification

Artificial Neural Networks (ANNs) and Spiking Neural Networks (SNNs) imitate the nervous system’s design of neurons (nodes) communicating through axons (edges).

While SNNs improve upon ANNs by incorporating time dynamics in the model’s inputs, neither method captures the time dynamics of signals between neurons in the network.

These dynamics have been shown to encode information in the brain, such as in audio processing. To capture these dynamics, we introduce the Digital Organoid, a spherical graph network that models the signals between neurons as traveling waves. The name and shape of the system take inspiration from brain organoids: neural cell cultures used to study neural development. Each axon in the network has electrophysiological properties determined by the physical dimensions of the system. These properties govern the behaviour of the signal wave along each axon. The network receives input stimuli using UV Mapping to project an image onto the surface of the sphere. The model is trained via a reinforcement learning agent that modifies the positions of nodes in the network to alter the speed of signals along each axon. The Digital Organoid expands upon SNNs to capture the dynamics of neural signalling over time and distance, as determined by each axon’s physiological properties, while learning is driven by conformational changes in the network’s structure.

Join at: imt.lu/conference