# Concept Bottleneck Models

Concept Bottleneck Models are neural networks that predict outcomes through explicit, human-interpretable intermediate concepts, enabling transparency and intervention. They differ from standard black-box models by constraining the decision pathway through understandable features.

Concept Bottleneck Models (CBMs) are a class of [neural network](https://www.wikiprompt.org/wiki/neural-network) architectures designed to make predictions in a two-stage process, with an explicit, human-interpretable intermediate layer of concepts. Instead of mapping raw inputs directly to outputs, a CBM first predicts a set of predefined, semantically meaningful attributes (the concepts), and then uses these predicted concepts to make the final decision. This design allows humans to inspect, understand, and even intervene on the model's internal reasoning, addressing a key limitation of standard black-box models in [[machine-learning|machine learning] and [deep learning](https://www.wikiprompt.org/wiki/deep-learning).

The concept bottleneck approach was formally introduced in a 2020 paper by Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Brandon Ying, and Percy Liang, affiliated with [Stanford University](https://www.wikiprompt.org/wiki/stanford-ai-lab) and [the University of Toronto](https://www.wikiprompt.org/wiki/university-of-toronto). The paper, titled "Concept Bottleneck Models," demonstrated that by forcing the model to predict interpretable concepts before the final output, one could achieve competitive accuracy while gaining the ability to debug and control the model's behavior. This work built on earlier ideas of interpretable and disentangled representations, but provided a concrete, trainable architecture that has since influenced a wide body of research.

## Architecture and Training

A Concept Bottleneck Model consists of three main components: an input encoder, a concept predictor, and a task predictor. The input encoder processes the raw data (e.g., an image, a text sequence, or a tabular record) into a feature representation. The concept predictor then maps this representation to a vector of concept scores, one for each predefined concept. Finally, the task predictor takes these concept scores (and optionally the original features) as input to produce the final output, such as a class label or a regression value.

Training a CBM typically follows one of three paradigms: independent training, sequential training, and joint training. In independent training, the concept predictor is first trained to predict ground-truth concept labels, and then the task predictor is trained on the predicted concepts (or on ground-truth concepts, with the concept predictor frozen). Sequential training trains the concept predictor first, then the task predictor while keeping the concept predictor fixed, but uses the predicted concepts during task training. Joint training optimizes both components simultaneously, often with a weighted loss that combines concept prediction error and task prediction error. Joint training usually yields the best task accuracy, but may sacrifice concept accuracy. A common variant is the "intervention-aware" training, where the model is trained to be robust to human corrections of concept predictions, allowing users to fix mistakes at test time.

## Interpretability and Intervention

The primary motivation for CBMs is interpretability. By constraining the model to reason through human-understandable concepts, users can see which concepts the model believes are present and how those concepts influence the final prediction. For example, in a bird classification task, a CBM might predict concepts like "has yellow belly," "has pointed wings," and "has a long tail," and then use those to classify the species. This transparency allows domain experts to verify that the model is using sensible features rather than spurious correlations.

A distinctive feature of CBMs is the ability to intervene. If a user knows that the model's concept prediction is wrong (e.g., it failed to detect a yellow belly), they can manually correct that concept value, and the task predictor will produce a new output based on the corrected concept. This human-in-the-loop capability is valuable in high-stakes domains like medicine or finance, where a model's mistake can be corrected before a final decision is made. Research has shown that even a small number of interventions can significantly improve accuracy, especially when the concept predictor is imperfect.

## Applications

CBMs have been applied across various domains. In medical imaging, they have been used for chest X-ray diagnosis, where concepts correspond to visible pathologies like "cardiomegaly" or "effusion." In dermatology, CBMs can predict skin lesion attributes before classifying malignancy. In autonomous driving, concepts like "pedestrian present" or "traffic light color" can be used to make driving decisions, though this is less common than in perception tasks. In natural language processing, CBMs have been applied to sentiment analysis and toxicity detection, with concepts such as "contains profanity" or "expresses anger." In tabular data, CBMs have been used for credit scoring, where concepts like "income level" and "payment history" are interpretable.

## Variants and Extensions

Several variants of CBMs have been proposed to address limitations. Label-free CBMs (introduced in 2022 by Tu, Ma, and others) remove the need for manual concept annotations by using a large pre-trained model (e.g., a [large language model](https://www.wikiprompt.org/wiki/large-language-model) or a vision-language model) to automatically generate concept labels from class descriptions. This reduces the annotation burden, which is a major bottleneck for applying CBMs to new tasks. Another variant, the Concept Embedding Model, projects concepts into a continuous embedding space rather than using binary or categorical labels, allowing for richer concept representations. Post-hoc CBMs attempt to extract a concept bottleneck from an already-trained black-box model, using probing or clustering techniques to identify interpretable dimensions.

Other extensions include probabilistic CBMs, which model concept uncertainty, and interactive CBMs, which allow users to query the model about which concepts are most influential. Some works have integrated CBMs with [transformer](https://www.wikiprompt.org/wiki/transformer) architectures, using attention mechanisms to align concepts with input regions.

## Limitations and Challenges

Despite their benefits, CBMs face several challenges. The most significant is the requirement for concept annotations, which are expensive and often subjective. Defining a complete and non-overlapping set of concepts for a task is difficult, and missing concepts can lead to information loss and reduced accuracy. CBMs also tend to underperform compared to standard black-box models on complex tasks, as the concept bottleneck restricts the information flow. The trade-off between interpretability and accuracy is a central research question.

Another issue is concept leakage: the task predictor might learn to use concept predictions in unintended ways, or the concept predictor might encode information beyond the stated concepts (e.g., through correlated features). This can make the model less interpretable than intended. Additionally, human interventions are only effective if the task predictor is robust to concept corrections; otherwise, correcting a concept might not change the output as expected.

## Relation to Other Interpretability Methods

CBMs are part of a broader family of interpretable machine learning techniques. They are often contrasted with post-hoc explanation methods like saliency maps or [gradient-based](https://www.wikiprompt.org/wiki/gradient-clipping) attribution, which attempt to explain a black-box model after training. CBMs are inherently interpretable by design, which is generally preferred for high-stakes decisions. They are also related to concept-based explanations, such as Testing with Concept Activation Vectors (TCAV), which use directional derivatives to measure concept influence, but TCAV does not constrain the model's architecture. CBMs are sometimes compared to decision trees and rule-based models, which are also inherently interpretable but less flexible.

## Current Research and Future Directions

Research on CBMs continues to evolve. Recent work has focused on scaling CBMs to large-scale datasets and integrating them with foundation models. For instance, using [CLIP](https://www.wikiprompt.org/wiki/clip)-style embeddings to define concepts in a continuous space has shown promise. Another direction is the use of CBMs in [reinforcement learning](https://www.wikiprompt.org/wiki/reinforcement-learning) and robotics, where interpretable intermediate representations can help with safety and debugging. There is also interest in using CBMs for fairness and bias mitigation, as concept bottlenecks allow auditors to check whether protected attributes are being used.

As of 2025, CBMs remain an active area of research, with papers appearing at major conferences like NeurIPS, ICML, and ICLR. The trade-off between interpretability and performance is still not fully resolved, but the field is moving toward more flexible and scalable concept-based models, including those that learn concepts dynamically rather than relying on fixed human-defined sets.

## See Also

- [interpretable-machine-learning](https://www.wikiprompt.org/wiki/interpretable-machine-learning)
- [explainable-ai](https://www.wikiprompt.org/wiki/explainable-ai)
- [neural-network](https://www.wikiprompt.org/wiki/neural-network)
- [deep-learning](https://www.wikiprompt.org/wiki/deep-learning)
- [machine-learning](https://www.wikiprompt.org/wiki/machine-learning)

---
Source: https://www.wikiprompt.org/wiki/concept-bottleneck-models
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-12T22:26:33.877282+00:00