(1) Past years witnessed the rise of Multimodal Foundation Models (MFM) trained on enormous datasets that can encode and process world information into feature representation. It is worth exploring how to retrieve informative knowledge from MFM, innovate their generative paradigms, and open up new research avenues for downstream AI tasks.
(2) Rapid growth of dataset scale and optimization complexity makes it hard to publicize vision toolkits to the greater community (e.g., laymen and non-experts). I envision significant importance on distilling pivotal knowledge from large-scale datasets to shed light on improving model trustworthiness and generalization.
(3) Generative models, especially the recent trends on multimodality, has triggered waves of revolution in the multimedia industry and empowered the deep learning community for various downstream tasks. Such potential intrigues me to investigate the intrinsic expressivity, sample efficiency, and controllability of generative models to provide constructive insights on high-quality, comprehensive, and responsible data synthesis.