Language in a Bottle: Language Model Guided Concept Bottlenecks for Interpretable Image Classification (2211.11158v2)

Abstract: Concept Bottleneck Models (CBM) are inherently interpretable models that factor model decisions into human-readable concepts. They allow people to easily understand why a model is failing, a critical feature for high-stakes applications. CBMs require manually specified concepts and often under-perform their black box counterparts, preventing their broad adoption. We address these shortcomings and are first to show how to construct high-performance CBMs without manual specification of similar accuracy to black box models. Our approach, Language Guided Bottlenecks (LaBo), leverages a LLM, GPT-3, to define a large space of possible bottlenecks. Given a problem domain, LaBo uses GPT-3 to produce factual sentences about categories to form candidate concepts. LaBo efficiently searches possible bottlenecks through a novel submodular utility that promotes the selection of discriminative and diverse information. Ultimately, GPT-3's sentential concepts can be aligned to images using CLIP, to form a bottleneck layer. Experiments demonstrate that LaBo is a highly effective prior for concepts important to visual recognition. In the evaluation with 11 diverse datasets, LaBo bottlenecks excel at few-shot classification: they are 11.7% more accurate than black box linear probes at 1 shot and comparable with more data. Overall, LaBo demonstrates that inherently interpretable models can be widely applied at similar, or better, performance than black box approaches.

• The paper introduces LaBo, leveraging GPT-3 to dynamically generate diverse concept candidates for improved image classification interpretability.
• It employs a novel submodular criterion for concept selection, achieving up to an 11.7% accuracy boost in few-shot scenarios.
• Human evaluations confirm that LaBo-derived concepts are more factual and visually grounded than those from traditional sources like WordNet.

Enhancing Interpretable Image Classification with LLM Guided Concept Bottlenecks

Introduction

The conventional method for enhancing the interpretability of deep learning models in high-stakes applications involves employing Concept Bottleneck Models (CBMs). These models interpret decisions through human-understandable concepts before arriving at a conclusion. However, the static nature and manual specification of concepts in CBMs hinder their widespread adoption due to their underperformance compared to their black-box counterparts, coupled with the extensive effort required for concept annotation. Addressing these limitations, the paper presents an innovative approach, named Language in a Bottle (LaBo), which leverages LLMs, specifically GPT-3, to dynamically generate and refine concept candidates, thereby providing an interpretable and accurate alternative to conventional models without the need for manually curated concepts.

Methodology

LaBo operates by prompting GPT-3 to generate descriptive sentences about image categories, which form candidates for concepts. These concepts are then refined through a novel selection process utilizing a submodular function emphasizing discriminability and diversity. This process effectively captures a wide range of informative and non-repetitive attributes. By aligning these textual concepts with image data using CLIP, a language-vision model, LaBo constructs a concept bottleneck that serves as a highly effective prior for critical concepts in visual recognition tasks.

The efficacy of LaBo is validated across 11 diverse datasets, showcasing its superiority in few-shot classification settings compared to linear probes and its competitive performance in larger data regimes. This strength is partly attributed to the method's ability to harness the rich world knowledge embedded in GPT-3, which, when coupled with a submodular optimization criterion, ensures the selection of informative and diverse concepts. Subsequent evaluation involving human annotators further confirms the interpretability and factual grounding of the generated concepts compared to those derived from existing baselines like WordNet or Wikipedia sentences.

Experimental Insights

Conducted experiments highlight several key findings:

• In few-shot scenarios, LaBo models surpass the accuracy of black-box linear probes by a significant margin (up to 11.7%) and remain competitive or outperform these models as the amount of data increases.
• The application of a novel submodular criterion is vital for selecting informative bottlenecks, underscoring the importance of discriminability and diversity in the concept selection process.
• Human evaluation metrics affirm that concepts generated by LaBo are more factual and visual compared to those sourced from WordNet or Wikipedia, thereby enhancing model interpretability without sacrificing accuracy.

Theoretical and Practical Implications

The proposed approach offers both theoretical and practical contributions to the field of interpretable AI. Theoretically, it demonstrates the possibility of constructing high-performing, interpretable models by leveraging external linguistic resources, without relying on manually annotated concepts or compromising on model transparency. Practically, LaBo provides a scalable framework for enhancing the interpretability of image classification models across various domains, potentially increasing trust in AI applications by enabling a clear understanding of model decisions.

Future Directions

Looking ahead, the paper opens avenues for further research in automating concept generation for interpretability. Future work might explore:

• Customized prompting strategies to improve concept relevancy and accuracy.
• Extension of LaBo to other tasks beyond classification, such as detection or segmentation.
• Investigation into the integration of LaBo with other LLMs and language-vision models to further enhance model interpretability and performance.

Conclusion

LaBo represents a significant advancement in leveraging natural language for improving the interpretability and accuracy of AI models. By dynamically generating and selecting concepts from LLMs, it circumvents the limitations associated with manual concept annotation in CBMs, offering a scalable and effective method for constructing interpretable AI systems.