Efficient Neural Architecture Search with Budget Constraints and Multi-Scale Entropy

After mutation, if the inference cost of the new structure (such as FLOPs, parameters, and latency) does not exceed the budget (such as budget L), and the depth is less than budget L, then Ft is added to the population p. The maximum depth L prevents the algorithm from generating structures that are too deep, which can result in high entropy and reduced performance. During the EA iteration, the overall size is kept constant by discarding the worst candidates with the minimum multi-scale entropy. At the end of evolution, the backbone with the highest multi-scale entropy is returned.