[Paper Review] Is Disentanglement all you need? Comparing Concept-based & Disentanglement Approaches

Dmitry Kazhdan, Botty Dimanov+4 more|arXiv (Cornell University)|Apr 14, 2021

Explainable Artificial Intelligence (XAI)14 references4 citations

TL;DR

This paper compares concept-based explanations and disentanglement learning approaches for interpretable AI, demonstrating that both state-of-the-art methods suffer from data inefficiency, task sensitivity, and vulnerability to concept representation choices. Despite shared goals in interpretability, neither approach is universally robust, highlighting critical trade-offs in design and training.

ABSTRACT

Concept-based explanations have emerged as a popular way of extracting human-interpretable representations from deep discriminative models. At the same time, the disentanglement learning literature has focused on extracting similar representations in an unsupervised or weakly-supervised way, using deep generative models. Despite the overlapping goals and potential synergies, to our knowledge, there has not yet been a systematic comparison of the limitations and trade-offs between concept-based explanations and disentanglement approaches. In this paper, we give an overview of these fields, comparing and contrasting their properties and behaviours on a diverse set of tasks, and highlighting their potential strengths and limitations. In particular, we demonstrate that state-of-the-art approaches from both classes can be data inefficient, sensitive to the specific nature of the classification/regression task, or sensitive to the employed concept representation.

Motivation & Objective

• To systematically compare concept-based explanations and disentanglement learning approaches in terms of their strengths, limitations, and underlying assumptions.
• To investigate how data efficiency, task dependence, and concept representation variance affect model performance and interpretability.
• To identify failure modes related to training data, labeling quality, and objective function design in both paradigms.
• To provide a unified benchmark library for future research on interpretable representation learning.
• To highlight that disentanglement alone does not ensure interpretability, and concept-based methods are not inherently robust to concept representation choices.

Proposed method

• Proposes a comprehensive library implementing state-of-the-art concept-based (CBM, CME) and disentanglement (VAE-based) models for comparative evaluation.
• Employs controlled experiments on dSprites and shapes3d datasets with varying concept dependencies and variance levels to isolate failure modes.
• Uses concept-to-task dependence analysis to measure how well concepts predict downstream labels across different tasks.
• Evaluates data efficiency by training models on progressively smaller subsets of data to assess performance drop-off.
• Tests concept variance fragility by manipulating input-level loudness of concepts (e.g., spatial or color differences) while keeping concept values constant.
• Applies both supervised (CBM) and post-hoc (CME) concept extraction, and compares weakly-supervised and unsupervised disentanglement methods.

Experimental results

Research questions

• RQ1 How do concept-based and disentanglement approaches compare in terms of data efficiency across diverse tasks?
• RQ2 To what extent is model performance sensitive to the specific nature of the downstream classification or regression task?
• RQ3 How does the choice and representation of concepts (e.g., color or spatial variance) affect model robustness and interpretability?
• RQ4 What are the failure modes of state-of-the-art models when concept representations are noisy or ambiguous?
• RQ5 Can disentanglement learning alone ensure human-interpretable representations, or are additional constraints required?

Key findings

• State-of-the-art concept-based and disentanglement models are both data inefficient, showing significant performance drops when trained on small datasets.
• Model performance is highly sensitive to the nature of the downstream task, with performance degrading when tasks depend on multiple or ambiguous concepts.
• Concept variance fragility is a major issue: even when concept values are identical, differences in input-level loudness (e.g., color distance in RGB space) significantly affect model performance.
• CBMs fail when the task relies on information not captured by provided concepts, leading to concept representations with 'impurities' that mix non-conceptual information.
• CME methods depend heavily on the quality of the pre-trained feature extractor φ, and their reliability drops if the hidden representation is not informative for concept prediction.
• Unsupervised disentanglement methods can learn infinitely many latent solutions, but none are guaranteed to be interpretable or useful without explicit supervision.