PAPER 9
Weakly supervised causal representation learning
Johann Brehmer (Qualcomm AI Research)
Pim de Haan (Qualcomm AI Research)
Phillip Lippe (formerly Qualcomm AI Research))
Taco Cohen (formerly Qualcomm AI Research)
NeurIPS 2022
Summary
In this paper we study under which conditions we can learn the “correct” high-level representations in embodied AI – meaning the representations associated with the data-generating process. We also provide a simple proof-of-concept example where we learn a representation that disentangles various buttons and lights from visual input, and then figures out which buttons a robot arm should press to cause different lights to turn on. We also introduce a new dataset which we call CausalCircuit. This allows us to explore causal representation learning in an intuitively causal setting. The CausalCircuit system consists of a robot arm that can interact with multiple touch-sensitive lights. The lights are connected through a stochastic circuit: a light is more likely to be on if its button is pressed or if its parent lights are on. The robot arm itself can be seen as part of the causal system.
Citation
@inproceedings{brehmer2022weakly, title = {Weakly supervised causal representation learning}, author = {Brehmer, Johann and De Haan, Pim and Lippe, Phillip and Cohen, Taco}, booktitle = {Advances in Neural Information Processing Systems}, year = {2022}, volume = {35}, eprint = {2203.16437}, url = {https://arxiv.org/abs/2203.16437}, }
Results
We introduced implicit latent causal models (ILCMs), which parameterize causal structure without requiring an explicit graph representation. We also discussed two algorithms for extracting the learned causal mechanisms and graph after training. In first experiments, we demonstrated that ILCMs let us reliably disentangle causal factors, identify causal graphs, and infer interventions from unstructured pixel data.
Looking for more papers with code?
* Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc.
