Dear all / Bonjour à tous,

Jury

Abstract

Most of machine learning consists of learning correlations in observed data. In causality, causal mechanisms are learned, which remains invariant to changes to other parts of the data-generating process, giving rise to out-of-distribution generalization guarantees. This is a desirable property for models learned using machine learning.

It is well known in the field of causality, interventions are generally necessary to identify the causal relationship between covariates unless we make strong assumptions on the data-generating process.
However, intervention-based strategies to identify causal relations are challenging to apply to several modalities of data in the real-world: data from biological experiments like high-throughput assays, medical images, natural language, etc. For these high-dimensional and unstructured observations, standard causal inference techniques face two challenges:
1. Intervening on every single measured variable often does not scale.
2. The causally relevant variables are not given, and thus must be inferred from the data.

Machine learning has been demonstrated to be very successful in settings when the data is high-dimensional. In-order to leverage the full capacities of both the field of causality which is concerned with learning causal relations, and machine learning whose successes has been attributed to being able to deal with high dimensional data, the two fields must be bridged by learning how to intervene on the world when presented with noisy, high dimensional data. Active learning is a sub-field of machine learning, where an algorithm intelligently requests labeled data from an oracle. Since interacting with the world to select data is baked into the design of active learning, we believe this existing sub-field of machine learning might help to give insights on how to design interventions to build causal models of the world.

In this proposal, we will explore how active learning can help to bridge this gap between machine learning and causality. This proposal consists of three new ideas, first the use of GFlowNets to model the posterior probability over DAGs and causal mechanisms, secondly, an insight into a potential new use-case of active learning, as a way to combat dataset bias, and finally, to propose a future research idea for query synthesis for active learning using disentangled representations.