We introduce the problem of hierarchical causal representation learning and show how our method, HERCULES, can generalize across environments with much fewer fine-grained interventions.
PDE-Refiner is an iterative refinement process that enables neural operator training for accurate and stable predictions over long time horizons. Published at NeurIPS 2023 (**Spotlight**).
BISCUIT identifies causal variables from videos using binary interactions between an external system (e.g. robot) and the causal variables. Published at UAI 2023 (**Spotlight**).
We present iCITRIS, a causal representation learning method that can identify causal variables with instantaneous effects and their graph from temporal sequences. Published at ICLR 2023.
We learn object-centric representations in an unsupervised manner by differentiably parameterizing and solving a min-cut problem to partition the image. Published at ICLR 2023.
We derive a minimal bound of experiments that guarantee identifiability of causal variables, opening up new opportunities for using intervention design for causal representation learning.
We present CITRIS, a causal representation learning algorithm for multidimensional causal factors identified from videos with interventions. Published at ICML 2022 (**Spotlight**).