Model Zoo

A "model zoo" refers to a collection of pre-trained deep learning models, readily available for researchers and practitioners to adapt for new tasks. Current research focuses on efficiently selecting, adapting, and ensembling models from these zoos, often employing techniques like reinforcement learning for data augmentation, graph learning for model selection, and novel methods for estimating model transferability across domains. This facilitates faster development and improved performance in various applications, ranging from wildlife monitoring and time series forecasting to visual recognition and personalized federated learning, while also enabling deeper investigation into the underlying properties and relationships between different model architectures.