Regularized Evolution for Image Classifier Architecture Search
share
The effort devoted to hand-crafting image classifiers has motivated the use of architecture search to discover them automatically. Reinforcement learning and evolution have both shown promise for this purpose. This study introduces a regularized version of a popular asynchronous evolutionary algorithm. We rigorously compare it to the non-regularized form and to a highly-successful reinforcement learning baseline. Using the same hardware, compute effort and neural network training code, we conduct repeated experiments side-by-side, exploring different datasets, search spaces and scales. We show regularized evolution consistently produces models with similar or higher accuracy, across a variety of contexts without need for re-tuning parameters. In addition, regularized evolution exhibits considerably better performance than reinforcement learning at early search stages, suggesting it may be the better choice when fewer compute resources are available. This constitutes the first controlled comparison of the two search algorithms in this context. Finally, we present new architectures discovered with regularized evolution that we nickname AmoebaNets. These models set a new state of the art for CIFAR-10 (mean test error = 2.13 parameters), and reach the current state of the art for ImageNet (top-5 accuracy = 96.2
READ FULL TEXT