Causal Concept Identification in Open World Environments

Moritz Willig, Matej Zecevic, Jonas Seng, Florian Peter Busch

Abstract

The ability to continually discover novel concepts is a core task in open world learning. For classical learning tasks, new samples might be identified via manual labeling. Since this is a labor-intensive task, this paper proposes to utilize causal information for doing so. Image data provides us with the ability to directly observe the physical, real-world appearance of concepts. However, the information presented in images is usually of a noisy and unstructured nature. In this position paper, we propose to leverage causal information to both structure and causally connect visual representations. Specifically, we discuss the possibilities of using causal models as a knowledge source for identifying novel concepts in the visual domain.

The Supervision Bottleneck in Open-World Learning

Modern computer vision relies heavily on supervised learning, where models are trained on massive, static datasets like ImageNet or MS COCO. However, this approach faces a significant hurdle in "open world" environments where new concepts appear constantly. The authors identify two major gaps in current methodologies:

  1. The Supervision Bottleneck: Manually labeling every new niche concept as it appears is unsustainable and labor-intensive.
  2. Lack of "Curiosity": Unlike humans, who use prior knowledge to infer what a new object might be based on its context and interactions, standard machine learning models treat every new concept as a blank slate. They fail to leverage the underlying "rules" of the physical world to aid discovery.

Methodology: The Causal Discovery Loop

The paper proposes a framework for Causal Concept Discovery based on Pearlian causality. The core idea is to use a "knowledge base" of causal structures to guide visual learning. The process involves several key components:

  • Structural Causal Models (SCMs): World knowledge is represented as SCMs—mathematical descriptions of how variables (concepts) relate to one another through functional equations.
  • The Causal Frontier: This represents the set of SCMs containing at least one concept the model already recognizes. By monitoring this "frontier," the model can predict which unknown concepts it is likely to encounter next.
  • Bootstrapping: The system begins with a small set of "known" concepts ($C_0$) identified via standard pre-trained models.

The Discovery Process

As illustrated in the match-flame example, the discovery loop follows four distinct stages:

  1. Identification: The system identifies known concepts in an image (e.g., a match and a matchbox).
  2. Selection: It searches its causal knowledge base for an SCM involving these known concepts (e.g., $Match + Matchbox \to Flame$).
  3. Inference: If the SCM predicts an effect (the flame) that the visual classifier does not yet recognize, the system uses the SCM's logic to "label" that specific region of the image.
  4. Retraining: The classifier is then retrained on this newly discovered concept, expanding its vocabulary autonomously.
Figure 1: Conceptual framework for identifying a novel 'Flame' concept using known 'Match' and 'Matchbox' concepts via a Structural Causal Model.

Key Insights

Causality as Supervision

The most provocative idea presented is that the structure of the physical world can replace the human in the loop. If the system understands that $A + B \to C$, and it observes $A$ and $B$ interacting, it receives a "free" supervision signal for $C$.

Disentanglement and Active Discovery

Unlike simple correlations—where a model might confuse a "match" with a "flame" because they often appear together—causal models help disentangle these concepts by defining the specific functional conditions under which a concept must appear. By defining a "Causal Frontier," researchers can move from passive learning to active discovery, where a model specifically "looks" for interactions to expand its vocabulary efficiently.

Limitations and Future Work

As a theoretical position paper, the authors acknowledge several significant hurdles:

  • The Abstraction Gap: Causal models usually operate at a high level (e.g., "Car"), while vision models operate at a low level (pixels). Bridging these—finding the "intermediate" concepts—remains a major challenge.
  • Identifying Causal Paths: In the real world, the path from cause to effect is often messy or obscured by high-dimensional pixel data.
  • The Inverse Problem: While this paper focuses on using causal knowledge to find visual concepts, the reverse—learning causal rules directly from open-world visual observations—is an equally necessary next step for autonomous AI.

Summary updated

Abstract

The ability to continually discover novel concepts is a core task in open world learning. For classical learning tasks new samples might be identified via manual labeling. Since this is a labor intensive task, this paper proposes to utilize causal information for doing so. Image data provides us with the ability to directly observe the physical, real-world appearance of concepts. However, the information presented in images is usually of noisy and unstructured nature. In this position paper we propose to leverage causal information to both structure and causally connect visual representations. Specifically, we discuss the possibilities of using causal models as a knowledge source for identifying novel concepts in the visual domain.
BibTeX
@inproceedings{willig2023causal,
  author = {Moritz Willig and Matej Zecevic and Jonas Seng and Florian Peter Busch},
  title = {{Causal Concept Identification in Open World Environments}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  url = {https://openreview.net/forum?id=tFpzHcJCoO},
}
Figure