Identifiable Representation Learning

Representation learning is at the heart of a paradigm shift in machine learning, driving many of today’s most advanced artificial intelligence (AI) systems. Representation learning transforms raw data into meaningful features that machines can use to perform a wide range of tasks, such as image recognition, content recommendation, and text generation. However, a key challenge remains: ensuring that these representations are identifiable, a property that is fundamental to reliably deploying AI systems in real-world applications. Although identifiability is critical to developing trustworthy AI systems, the current mathematical understanding of representation learning in general, and identifiability in particular, is extremely limited. To address this gap in understanding, Banff International Research Station will host a workshop focusing on principled approaches for learning identifiable representations and understanding the properties of such representations.