AI for Pattern Discovery can meet the needs of Environmental Science

Abstract:

Today, in the environmental sciences, the means by which we obtain hypotheses or models is taken for granted: it is human intuition. In contrast, while modern AI algorithms enable astounding predictive performance in some of the world’s most complex systems – they fall woefully short of imparting human understanding. It is time to bring to bear the might of modern mathematics to overcome the enormous theoretical challenges that stand between the state of the art and our capacity to extract internal data representations from fitted AI learners. Ultimately, AI algorithms must be tasked as we now uniquely task human intuition: we must develop methods to extract novel, testable hypotheses directly from learning machines. Here, we present recent advances in theory and algorithms that have significantly improved the interpretability and explainability of AI algorithms. We describe a framework for “Pattern Discovery” and demonstrate its utility in the derivation of “first-principles” models directly from data. We describe success stories in the environmental and agricultural sciences -- where reduced order, especially monoculture ecosystems, including microbiomes and environmental variables enable exquisite control of system parameters. We demonstrate the discovery of ecosystem control points from data, and their efficacy for obtaining predictive understanding of environmental systems. We point to a future for environmental data science defined by our capacity to learn from learning machines.