Emergent Mind

Where is the Truth? The Risk of Getting Confounded in a Continual World

Published Feb 9, 2024 in cs.LG and stat.ML


A dataset is confounded if it is most easily solved via a spurious correlation which fails to generalize to new data. We will show that, in a continual learning setting where confounders may vary in time across tasks, the resulting challenge far exceeds the standard forgetting problem normally considered. In particular, we derive mathematically the effect of such confounders on the space of valid joint solutions to sets of confounded tasks. Interestingly, our theory predicts that for many such continual datasets, spurious correlations are easily ignored when the tasks are trained on jointly, but it is far harder to avoid confounding when they are considered sequentially. We construct such a dataset and demonstrate empirically that standard continual learning methods fail to ignore confounders, while training jointly on all tasks is successful. Our continually confounded dataset, ConCon, is based on CLEVR images and demonstrates the need for continual learning methods with more robust behavior with respect to confounding.


  • Introduces the ConCon dataset built for studying confounding in continual learning (CL) environments, addressing challenges in generalizing learning across tasks.

  • Explores the vulnerability of common CL methods like experience replay and elastic weight consolidation to 'insidious continual confounding' in environments with spurious correlations.

  • Emphasizes the importance of distinguishing between genuine and spurious correlations in CL to build models that generalize well across sequential tasks.

  • Suggests future research directions, including the integration of causal reasoning in CL methods and further exploration of the unique challenges posed by the ConCon dataset variants.

In the fast-evolving domain of continual learning (CL), a new dataset, named ConCon, emerges to tackle the often overlooked challenge of confounding in continually evolving environments. This recent contribution by Florian Peter Busch et al. introduces ConCon, a synthetic dataset built on the CLEVR framework, specifically designed for the systematic study of confounding in CL scenarios. The dataset is accompanied by a comprehensive exploration of how existing CL methods grapple with scenarios where models may latch onto spurious correlations that do not generalize across tasks—a phenomenon termed as “insidious continual confounding.”

The ConCon Dataset: A Brief Overview

ConCon operates on a simple premise: it consists of images of geometric objects, where the task is to classify these images based on a ground truth rule. This rule remains consistent across the dataset, but the introduction of confounding variables in a sequential task setup poses a unique challenge. These confounders are characteristics that might make the task easily solvable within a single task's context but jeopardize the model's ability to generalize this learning to future, unseen tasks.

The dataset splits into two variants: "disjoint," where confounders are isolated within their respective tasks, and "strict," where confounders, notwithstanding their relevance, may appear across tasks albeit being informative only within their specific task. Each variant proposes a different spectrum of challenge in identifying and adhering to the ground truth amidst the potential misguidance by confounders.

The Perils of Continual Confounding

Through a series of experiments utilizing common CL methods such as experience replay (ER) and elastic weight consolidation (EWC), alongside the examination on both neural network (NN) and neuro-symbolic (NeSy) models, the study reveals a significant vulnerability in current CL approaches. The findings indicate that while methods like ER can mitigate catastrophic forgetting, they fall short in circumventing the pitfalls of continual confounding. Particularly noteworthy is the emergence of "insidious continual confounding" in the strict setting, where CL methods underperform in comparison to a joint training scenario despite being exposed to the same data. This discrepancy underscores the difficulties CL models face in discerning and retaining the ground truth when sequentially exposed to confounded tasks.

Implications and the Path Forward

The revelations from the ConCon dataset highlight a crucial aspect of CL that extends beyond the traditional problem of forgetting—namely, the risk of learning incorrect or non-generalizable patterns due to confounding. The implications of these findings are twofold. Practically, it presents an immediate challenge to the deployment of CL systems in dynamic real-world settings, where the ability to discern and adapt to the fundamental underlying rules amidst changing conditions is imperative. Theoretically, it prompts a reconsideration of current CL methods and invites the development of novel strategies capable of distinguishing between spurious and genuine correlations, thereby learning a more robust and generalizable model.

Future Directions in CL Research

Looking ahead, the ConCon dataset not only offers a valuable tool for benchmarking and improving existing CL methodologies but also opens new avenues for research. Of particular interest might be the exploration into methods that incorporate causal reasoning to better understand and mitigate the effects of confounders. Additionally, the distinct challenges posed by the disjoint and strict variants of the dataset warrant further investigation into tailored approaches that can dynamically adjust to the nature of confounders encountered in a learning sequence.

In conclusion, the ConCon dataset serves as a critical reminder of the complexities inherent in CL and the importance of designing models that are not only resistant to forgetting but are also adept at navigating the labyrinth of continually evolving data landscapes without being misled by confounders. As the field of CL progresses, the lessons drawn from ConCon will undoubtedly play a pivotal role in shaping more resilient and intelligent learning systems.

Get summaries of trending AI/ML papers delivered straight to your inbox

Unsubscribe anytime.