The Department of Energy Office of Science (SC) program in Advanced Scientific Computing Research (ASCR) hereby announces its interest in making research data and artificial intelligence (AI) models findable, accessible, interoperable, and reusable (FAIR1) to facilitate the development of new AI applications in SC’s congressionally authorized mission space, which includes the advancement of AI research and development. In particular, ASCR is interested in supporting FAIR benchmark data for AI; and FAIR frameworks for relating data and AI models
For this FOA, AI is inclusive of, for example, machine learning (ML), deep learning (DL), neural networks (NN), computer vision, and natural language processing (NLP). Data, in this context, are the digital artifacts used to generate AI models and/or employed in combination with AI models during inference. An AI model is an inference method that can be used to perform a “task,” such as prediction, diagnosis, or classification. The model is developed using training data or other knowledge. An AI task is the inference activity performed by an artificially intelligent system.
Tools for training AI models on data are readily available and widely used. What is lacking, however, is a theoretical framework for understanding relationships between data and models. For example, given a specific data set and problem, we lack rigorous methods for identifying the best model, hyper-parameters, and training method to use. Given a specific data set and problem, which additional data would be helpful to include in the training set? What information about a dataset can be deduced from a model trained on the data? What attributes of the data can be reverse engineered from a model? What can we learn about model robustness and transfer learning by looking at relationships between data and models?
The primary focus of this FOA topic is to advance our understanding of the relationship between data and models by exploring relationships among them through the development of FAIR frameworks for relating data and models. Such frameworks should provide capabilities that advance our understanding of AI, provide new insights to help researchers with applications of AI techniques, and provide an environment where novel approaches to AI can be explored.
Proposed frameworks may focus on specific disciplines or sub-disciplines currently supported by SC’s programs in ASCR, Biological and Environmental Research (BER), Basic Energy Sciences (BES), Fusion Energy Sciences (FES), High Energy Physics (HEP), Nuclear Physics (NP), or may focus on particular aspects or sub-areas within AI.