LG-CAV: Train Any Concept Activation Vector with Language Guidance

Making AI explanations easier to create: Traditional methods for explaining AI decisions through "concept activation vectors" required expensive, manually-labeled datasets of thousands of images. LG-CAV eliminates this bottleneck by using descriptions of concepts (text) instead, allowing researchers to explain AI decisions for virtually any concept without costly data collection.

Leveraging existing AI knowledge: The method works by tapping into knowledge already embedded in pre-trained vision-language models like CLIP, rather than starting from scratch. This represents a practical approach to responsible AI by reusing existing resources to improve model interpretability across different AI systems.

Improving model accuracy as a bonus: Beyond just explaining decisions, the researchers demonstrated that their concept-based approach can actually improve the target model's performance—tested across four datasets and nine different neural network architectures. This suggests that understanding what concepts drive AI decisions can simultaneously make those decisions more reliable.

Scalability for real-world deployment: By removing the need for labeled training data, LG-CAV makes concept-based explanations practical for deployment in real applications where predefined concept sets may be insufficient or where new concepts emerge over time.