In the pursuit of interpretable deep learning, two primary paradigms have emerged: Concept-Based Explanations and Disentanglement Learning. While both aim to decompose complex data into human-understandable units, they have largely developed in isolation. This paper provides a systematic comparison between these approaches, evaluating their data efficiency, task dependency, and robustness to feature variance.
The Interpretability Gap
Deep learning models are often criticized for being "black boxes." Researchers have proposed two main solutions:
- Concept-Based Models (CBMs): These models are explicitly trained to identify human-defined concepts (e.g., "shape," "color") as an intermediate step before final prediction.
- Disentanglement (VAEs): Generative models, such as Variational Autoencoders, attempt to automatically discover independent latent factors of variation without explicit human labels.
This study investigates whether these methods are truly interchangeable and identifies the inherent trade-offs in each approach using benchmark datasets like dSprites and shapes3D.
Data Efficiency and the Labeling Burden
A significant hurdle for both paradigms is the requirement for annotated data. While unsupervised VAEs exist, achieving reliable interpretability often requires weak supervision (WVAEs). The authors find that both CBMs and WVAEs are surprisingly data-hungry.
As shown in Figure 1, CBMs generally demonstrate higher data efficiency than WVAEs. This is because CBMs are directly optimized for concept identification, whereas WVAEs must balance disentanglement with a reconstruction loss. To achieve even moderate concept accuracy (e.g., 60%), these models often require up to 40% of the dataset to be manually labeled.
Concept "Loudness" and Variance Bias
One of the paper's most critical insights is the concept of "loudness." Generative models like VAEs are naturally biased toward features that cause the highest variance in raw pixel values, such as object position or large-scale color changes. These are termed "loud" concepts.
In Figure 5, we observe that WVAEs struggle to capture "quiet" concepts (like subtle rotations) when "loud" concepts (like spatial position) are present. In contrast, CBMs remain robust to this variance because their supervision forces the model to attend to all defined concepts regardless of their pixel-level impact. This "variance fragility" suggests that unsupervised methods may systematically ignore subtle but important attributes.
Task Dependency and Extraction Limits
The study also examines Post-hoc Concept Extraction (CME), where researchers attempt to find concepts within a pre-trained model. The results indicate a "Task-Dependency Trap": if the original model was trained on a task that did not require a specific concept, that concept is effectively lost. You cannot extract information that the model never learned to represent. Accuracy in concept extraction drops significantly as the end-task becomes less dependent on those specific features.
Conclusion
The comparison reveals that "interpretable" does not always mean "reliable." While CBMs offer better data efficiency and robustness to feature variance, they require heavy supervision. Disentanglement methods offer a path toward discovery but are currently limited by a bias toward high-variance features. Future work must address these trade-offs to build models that are both interpretable and practical for complex, real-world applications.
Summary updated