Discovering Latent Structures from Data

It can be argued that intelligent agents such as humans are constantly building abstract models and using them to reason about the world around them. An important question in AI today is whether this process can be automated and whether such models can be useful in applications. In this talk, I will discuss three papers corresponding to three settings where this can be done with theoretical guarantees. Firstly, I’ll briefly mention how interpretable latent factors of variations can be discovered from data by learning a sparse latent causal graph relating them. Secondly, I will present a recent work exploring the usefulness of disentangled representations and how to learn such representations via sparse multitask learning. Thirdly, we will look into the problem of learning a causal graph from observations and see how this can be achieved via continuous-constrained optimization. Finally, I’ll present a unifying view of these various settings and use it to suggest future lines of investigation.