On Supernet Transfer Learning for Effective Task Adaptation

Neural Architecture Search (NAS) methods have been shown to outperform hand-designed models and help to democratize AI. However, NAS methods often start from scratch with each new task, making them computationally expensive and limiting their applicability. Transfer learning is a practical alternative with the rise of ever-larger pretrained models. However, it is also bound to the architecture of the pretrained model, which inhibits proper adaptation of the architecture to different tasks, leading to suboptimal (and excessively large) models. We address both challenges at once by introducing a novel and practical method to textittransfer supernets, which parameterize both weight and architecture priors, and efficiently finetune both to new tasks. This enables supernet transfer learning as a replacement for traditional transfer learning that also finetunes model architectures to new tasks. Through extensive experiments across multiple image classification tasks, we demonstrate that supernet transfer learning does not only drastically speed up the discovery of optimal models (3 to 5 times faster on average), but will also find better models than running NAS from scratch. The added model flexibility also increases the robustness of transfer learning, yielding positive transfer to even very different target datasets, especially with multi-dataset pretraining.