Events

CDS Seminar: From Distributed Representations to Causal Representation Learning

Speaker: Francesco Locatello

Location: On-Line
Videoconference link: https://nyu.zoom.us/j/91634861736?pwd=TW4rWmxzdXJMRzRyZ1FHdGdRaytQZz09

Date: Wednesday, March 9, 2022

Sitting at the intersection between machine learning and graphical causality, causal representation learning tackles the problem of discovering high-level variables from low-level observations. I will motivate why such abstractions may be useful for deep learning, covering case studies in zero-shot generalization of RL policies and pitfalls in training fair classifiers. The core of the talk will cover how to learn causal representations. I will address identifiability results for both disentanglement and causal representation learning, highlighting the practical challenges and the conceptual limitations of these frameworks. I will present a way forward through architectures adapting the computation in their forward pass. I will conclude presenting a new approach for causal discovery that leverage advances in score based generative models.