Human-centered artificial intelligence (AI) envisions a future where AI systems augment and cooperate with humans seamlessly. Towards this goal, AI systems, such as self-driving cars and artificial clinicians, need to imitate human expertise and self-adapt to novel situations or tasks. In this thesis, we propose machine learning methods to advance AI systems' abilities to imitate and adapt. The proposed methods combine principled theoretical motivation with deep learning's expressivity to obtain robust models with strong performances.

For imitation learning (IL), we present a novel strategy based on support estimation, which computes a fixed reward and recasts IL as standard reinforcement learning (RL). Our approach improves computational efficiency and learning stability over adversarial IL methods. We also unify support estimation with adversarial IL by imposing a soft consistency constraint between them. The resulting framework combines their respective strengths and improves algorithmic stability and performance.

Casting the problem of self-adaptation as meta-learning, we propose a novel meta-learning framework based on structured prediction. The proposed framework explicitly ranks past experiences with a new task and learns from the most relevant experiences. Our method can capture complex task distributions for improved generalization and offers a principled and interpretable mechanism for utilizing past experiences.

Beyond standard benchmarks, we apply the methods developed in this thesis on two challenging tasks from intelligent vehicles, including imitating human driving behaviors and personalized monitoring of drivers' cognitive load. Compared to benchmarks, our proposed methods demonstrate improved learning ability and generalization performance. We also discuss future directions for our research.