TL;DR
This paper introduces Concept-Centric Transformers (CCTs), a novel framework that enhances the interpretability of deep learning models by integrating object-centric concept learning within a shared global workspace. CCTs achieve state-of-the-art classification accuracy while providing more consistent and faithful concept-based explanations, making them suitable for safety-critical applications.
What Problem Does This Paper Address?
Modern deep learning models, despite their impressive performance, often suffer from a lack of transparency, making it difficult to understand why they make certain decisions. This "black box" nature limits their adoption in critical domains like healthcare and civil engineering, where trust and explainability are paramount. While existing methods offer post-hoc explanations (explaining a trained model after the fact) or focus on low-level features, this paper addresses the need for intrinsically interpretable models that generate human-understandable "concepts" as part of their decision-making process. The challenge lies in effectively configuring these interpretable models to facilitate communication and specialization among different computational modules, inspired by theories of cognitive architecture.
Methods
The core of the proposed solution is the Concept-Centric Transformer (CCT), which acts as a drop-in replacement for the classifier head of existing backbone models. CCT is built upon the Shared Global Workspace (SGW) theory, which posits that distributed modules can benefit from competing for access to a shared memory. CCT comprises three key components:
- Concept-Slot-Attention (CSA) Module: This module acts as the shared memory. It extracts semantic concepts from input features by learning "concept slots." It utilizes variants of Slot-Attention (SA), a technique for unsupervised object-centric learning, to process input embeddings and produce concept representations.
- Cross-Attention (CA) Module: This module facilitates communication. It generates classification outputs by performing cross-attention between the input features and the learned concept embeddings from the CSA module. This allows the model to relate concepts to the input data.
- Specialized Loss Functions:
- Explanation Loss: This loss encourages faithfulness by leveraging expert knowledge to guide attention heads towards specific concepts. It acts as a regularization term, penalizing deviations from a desired attention distribution.
- Sparsity Loss: This entropy-based loss encourages sparsity in attention, helping to determine feature importance during training.
The CCT architecture is designed to augment existing deep learning backbones, adding explainability without sacrificing performance.
Datasets
The paper evaluates CCT on three benchmark datasets:
- CIFAR100 Super-class: Used to demonstrate CCT's capabilities under fully supervised conditions with available global concept explanations.
- CUB-200-2011: Used to showcase performance in both supervised and unsupervised explanation setups.
- ImageNet: Used to challenge CCT in a fully unsupervised scenario, demonstrating its ability to learn concepts without explicit explanations.
Key Insights and Ideas
- SGW as a Framework for Interpretability: The paper successfully translates the cognitive theory of a Shared Global Workspace into a computational framework for building interpretable AI models. This provides a principled way to manage interactions between specialized modules.
- Concept Slots as a Form of Memory: The CSA module's "concept slots" act as a bottlenecked shared memory, forcing specialization and compositionality among learned concepts, leading to more robust representations.
- Faithfulness by Design: The Cross-Attention module is designed to ensure that explanations are faithful to the model's decision-making process, a critical aspect of interpretability.
- Unsupervised Concept Learning is Viable: The experiments on ImageNet demonstrate that CCT can learn meaningful and semantically coherent concepts even without explicit ground-truth explanations, highlighting its versatility.
- Concept Learning Enhances Classification: The paper shows that by learning interpretable concepts, CCT not only provides better explanations but also achieves superior classification accuracy compared to many existing methods.
Key Results
- State-of-the-Art Accuracy: CCT consistently outperforms all baselines on classification tasks across CIFAR100 Super-class, CUB-200-2011, and ImageNet. For instance, on CIFAR100 Super-class, CCT achieved 92.6-93.0% accuracy with ViT-T backbone, significantly outperforming CT (92.1%). On CUB-200-2011, CCT achieved over 90% accuracy with various backbones, surpassing non-interpretable baselines.
- More Consistent Explanations: Qualitative examples show that CCT generates more accurate and semantically aligned concept explanations compared to prior methods like CT, which sometimes hallucinate incorrect concepts.
- Effective Unsupervised Learning: On ImageNet, CCT achieved 83.7% accuracy with a small ViT-S backbone, meeting or exceeding other SOTA methods, even without concept explanations.
Limitations and Future Work
- Hyperparameter Tuning: While the paper mentions that some SA variants might depend on hyperparameter tuning, the CCT formulation aims to mitigate this. However, the general sensitivity to hyperparameters in deep learning models remains a consideration.
- Composable Knowledge: The authors suggest extending the concept extraction module to acquire "composable pieces" of knowledge and learning the rules that govern their composition. This could lead to even more sophisticated and generalizable explanations.
- Broader Applications: Future work could explore applications beyond image classification, such as model debugging and medical-image diagnosis, where interpretability is crucial.
Summary updated