MAE-NAS: Efficient Backbone Design with Multi-Scale Differential Entropy and Evolutionary Algorithm

MAE-NAS utilizes a customized evolutionary algorithm to detect the multi-scale differential entropy of the backbone. To enhance evolutionary efficiency, a coarse-to-fine strategy is employed to gradually reduce the search space. Initially, N network architectures are randomly generated to populate the population 'p'. A network architecture 'Ft' consists of a sequence of building blocks, such as ResNet blocks. Subsequently, a block is randomly chosen and replaced with its mutated counterpart. Coarse mutation is applied in the early stages, transitioning to fine mutation after T/2 iterations. During coarse mutation, block type, kernel size, depth, and width are randomly mutated. In fine mutation, only kernel size and width undergo mutation.