Continuous Ant-Based Neural Topology Search

Abstract

Ant-based Topology Search (ANTS) is a Neural Architecture Search (NAS) inspired by ant colony optimization (ACO). ANTS encodes the neural structure search space within a highly interconnected structure. Optimization agents, like ants, navigate this structure in search of an optimal neural topology. Continuous Ant-based Topology Search (CANTS) builds upon ANTS by replacing the discrete search space with a 3D continuous one. CANTS introduces a fourth dimension for potential neural synaptic weights, transitioning from NAS to NeuroEvolution (NE). This automates artificial neural network design without relying on backpropagation, reducing optimization time and offering a promising approach for machine learning applications.