Learning Latent Structural Causal Models
Jithendaraa Subramanian, Yashas Annadani, Ivaxi Sheth, Nan Rosemary, Ke, Tristan Deleu, Stefan Bauer, Derek Nowrouzezahrai, Samira Ebrahimi Kahou
TL;DR
This paper introduces a Bayesian inference method to learn latent structural causal models directly from low-level data like images, enabling causal understanding and out-of-distribution generalization.
Contribution
It presents a tractable approximate inference approach for jointly learning the structure, variables, and parameters of latent SCMs from low-level data with interventions.
Findings
Effective joint inference on synthetic data
Successful causal image generation from unseen interventions
Demonstrated out-of-distribution generalization
Abstract
Causal learning has long concerned itself with the accurate recovery of underlying causal mechanisms. Such causal modelling enables better explanations of out-of-distribution data. Prior works on causal learning assume that the high-level causal variables are given. However, in machine learning tasks, one often operates on low-level data like image pixels or high-dimensional vectors. In such settings, the entire Structural Causal Model (SCM) -- structure, parameters, \textit{and} high-level causal variables -- is unobserved and needs to be learnt from low-level data. We treat this problem as Bayesian inference of the latent SCM, given low-level data. For linear Gaussian additive noise SCMs, we present a tractable approximate inference method which performs joint inference over the causal variables, structure and parameters of the latent SCM from random, known interventions. Experiments…
Peer Reviews
Decision·Submitted to ICLR 2024
My primary concerns are as follows: 1) Theoretical guarantees: It is widely recognized that identifying latent causal models is a challenging task without the incorporation of additional assumptions. Recent studies have made significant progress in demonstrating the identifiability of latent causal models by exploring the change of weights, such as hard and soft interventions [1][2], on the model's weights. Nevertheless, there has been a noticeable absence of discussion regarding how our propos
See above
This paper is well written with clear motivation. What the authors focused on is indeed an interesting yet challenging research topic in causal inference and machine learning.
Novelty: In my opinion, the authors introduced an approach for parameter estimation through deep learning methods. However, it's worth noting that they didn't provide a theoretical analysis to support their approach. That is to say, the authors did not offer an analysis of the identifiability of the latent causal model. Without theoretical identifiability results, it becomes challenging to have full confidence in the outcomes generated by their proposed method. Experiments: The experimental res
This paper considers the estimation of latent causal models, which is a very important task in causal representation learning.
The novelty of this work appears somewhat constrained. It focuses solely on the scenario where the latent causal model is linear, interventions on latent variables are assumed, and the intervention targets are known. Furthermore, it does not explicitly clarify whether this setting is theoretically identifiable. The experimental validation is somewhat lacking. The paper only presents results with 5 latent variables, which is not enough for an empirical study.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsBayesian Modeling and Causal Inference · Domain Adaptation and Few-Shot Learning · Explainable Artificial Intelligence (XAI)