This paper focuses on the design of an architecture for efficient prior information acquisition in fields such as computational chemistry, immunology, and medical imaging, where large-scale prior learning models cannot be used due to lack of data. We demonstrate that neural network memory can be used to adapt to non-stationary distributions with a small number of samples, and that a hypernetwork trained with model-agnostic meta-learning (MAML) (a network that generates other networks) can acquire more generalized prior information than a standard network. By applying the hypernetwork to 3D scene generation, we achieve fast text-to-3D generation through efficient prior information acquisition with only a small number of training scenes, and perform 3D segmentation of new scenes with limited data. Finally, we reuse existing molecule generation methods as a pre-training framework to improve molecular feature prediction, which is a major challenge in computational immunology.
