Towards the Reusability and Compositionality of Causal Representations

TL;DR

This paper addresses the challenge of reusing and composing causal representations learned from one environment in new, related environments with interventions. It introduces DECAF, a framework that detects which causal factors must be adapted and which can be transferred, by exploiting intervention targets in the TempoRal Intervened Sequences (TRIS) setting and integrating with existing CRL methods. A simple generative model formalizes adaptation and composition of causal representations across environments, and DECAF is validated on three benchmarks, showing improved data efficiency and robust transfer when combined with CITRISVAE, LEAP, DMSVAE, and iVAE. The work demonstrates practical benefits for rapid adaptation and cross-domain composition, with potential impact on transfer learning, RL, and causal reasoning in high-dimensional observation spaces.

Abstract

Causal Representation Learning (CRL) aims at identifying high-level causal factors and their relationships from high-dimensional observations, e.g., images. While most CRL works focus on learning causal representations in a single environment, in this work we instead propose a first step towards learning causal representations from temporal sequences of images that can be adapted in a new environment, or composed across multiple related environments. In particular, we introduce DECAF, a framework that detects which causal factors can be reused and which need to be adapted from previously learned causal representations. Our approach is based on the availability of intervention targets, that indicate which variables are perturbed at each time step. Experiments on three benchmark datasets show that integrating our framework with four state-of-the-art CRL approaches leads to accurate representations in a new environment with only a few samples.

Figures (10)

• Figure 1: Overview of our approach for the adaptation task in Pong, where the source environment on which we learn the initial causal representation models the position of the ball as Cartesian coordinates, while the target environment uses polar coordinates for the ball position.
• Figure 2: Overview of our approach for the composition task in Pong, where the first source environment models the data with polar ball position and independent paddles and the second environment employs Cartesian ball coordinates but entangled paddles. The target environment uses Cartesian coordinates for the ball position and has independent paddles.
• Figure 3: Spearman CC (higher best, $\uparrow$) of inferred latents to the ground truth changed variables when adapting the representations. CRL approaches are color-coded, the proposed method has a darker color.
• Figure 5: Spearman CC (higher best, $\uparrow$) when adaptating and composing with increasing number of samples. Solid lines describe the mean while shaded areas the standard deviation over 5 runs. (a) Correlation of changed factors for DMSVAE approach when adapting from CA $\rightarrow$ PO in InterventionalPong. (b) Correlation of all factors when composing REG-j+CH-i$\rightarrow$REG-i in Voronoi Benchmark.
• Figure 6: Spearman CC (higher best, $\uparrow$) of inferred latents to the ground truth of all variables when composing representations. CRL approaches are color-coded, the proposed method has a darker color. (a) Composition of factors in InterventionalPong with sources CA-jPA and PO-PA. (b) Composition of factors in Causal3DIdent with sources CA-jHUE and ROT-HUE.