Metric Aware Abstraction

Metric-aware abstraction in artificial intelligence focuses on developing methods for automatically creating hierarchical representations of data and tasks, mirroring human cognitive abilities to simplify complex problems. Current research explores diverse approaches, including neuro-symbolic methods combining neural networks with symbolic reasoning, and the use of optimal transport and reinforcement learning algorithms to learn efficient abstractions from data. This research aims to improve the efficiency, robustness, and generalizability of AI systems across various domains, from reinforcement learning and machine learning to program synthesis and causal inference.