MINDS 2021 Winter Symposium- Luiz Chamon
Title– Learning under requirements
Abstract– The transformative power of learning lies in automating the design of complex systems, allowing us to go from data to operation with little to no human intervention. Today, however, learning today does not incorporate requirements organically, which has led to data-driven solutions prone to tampering, unsafe behavior, and biased, prejudiced actions. To realize its autonomous engineering potential, we must develop learning methods capable of satisfying requirements beyond the training data. In this talk, I will show when and how it is possible to learn under requirements by developing the theoretical underpinnings of constrained learning. I will define constrained learning by extending the classical probably approximately correct (PAC) framework and show that despite appearances, constrained learning is not harder than unconstrained learning. In fact, they have essentially the same sample complexity. I will also derive a practical learning rule that under mild conditions can tackle constrained learning tasks by solving only unconstrained empirical risk minimization (ERM) problems, a duality that holds despite the lack of convexity. I will illustrate how these advances enable the data-driven design of trustworthy systems that adhere to fairness, robustness, and safety specifications. I see these contributions as steps towards a constraint-driven learning paradigm that I will briefly discuss together with the new theoretical and practical questions it raises.