A Multi-objective Evolutionary Algorithm Based on Bi-population with Uniform Sampling for Neural Architecture Search
A clever two-part search strategy helps AI build more efficient and powerful neural networks.
Deep Dive
Researchers have developed a new AI method, MOEA-BUS, that automates the design of neural networks. It uses a two-population evolutionary search and a uniform sampling technique to simultaneously optimize for high accuracy and low computational complexity. The method achieved 98.39% accuracy on CIFAR-10 and 80.03% on ImageNet, while also finding a model with 78.28% accuracy using only 446 million multiply-add operations.
Why It Matters
This makes creating efficient, high-performance AI models faster and less reliant on expert human designers.