Scientists at the IT University of Copenhagen propose self-assembling neural networks, inspired by biological systems, as a novel approach in AI research. Unlike deep neural networks with predefined structures, these networks emulate the organic formation of biological neural networks, starting from a single neuron and evolving through a self-organizing process. The concept involves a graph neural network encoding with two interconnected networks: a policy network for agent actions and a Neural Developmental Program (NDP) controlling each neuron's growth and connections. The NDP allows autonomous, decentralized growth, similar to biological counterparts. Initial tests show promise in solving problems, albeit not as optimally as existing solutions. The approach's limitations include performance deterioration after a certain growth stage and a lack of activity-dependent growth. Researchers suggest exploring the interplay between genome size, developmental steps, and task performance.
Scientists at the IT University of Copenhagen propose self-assembling neural networks, inspired by biological systems, as a novel approach in AI research. Unlike deep neural networks with predefined structures, these networks emulate the organic formation of biological neural networks, starting from a single neuron and evolving through a self-organizing process. The concept involves a graph neural network encoding with two interconnected networks: a policy network for agent actions and a Neural Developmental Program (NDP) controlling each neuron's growth and connections. The NDP allows autonomous, decentralized growth, similar to biological counterparts. Initial tests show promise in solving problems, albeit not as optimally as existing solutions. The approach's limitations include performance deterioration after a certain growth stage and a lack of activity-dependent growth. Researchers suggest exploring the interplay between genome size, developmental steps, and task performance.