AI2030 Responsible AI Tools Marketplace

Key Takeaways

AI models can develop problematic "shortcuts" that undermine real-world performance: The research demonstrates that sonar image classifiers achieved high accuracy in lab settings but relied too heavily on seafloor features rather than learning to identify objects properly, illustrating how accuracy metrics alone don't guarantee reliable AI systems.
"Unlearning" is a practical approach to fixing AI bias after deployment: Rather than retraining models from scratch, the researchers developed a method to selectively remove learned biases from existing models, offering a more efficient way to correct AI systems that have learned to depend on irrelevant background features.
Explainability tools are essential for verifying that AI bias removal actually worked: The framework includes visualization methods that show what the model has "forgotten," allowing researchers and stakeholders to confirm that bias mitigation efforts were successful—not just assume improvements based on performance scores.
Responsible AI in high-stakes domains requires multiple safeguards: For applications like sonar detection in defense and civilian contexts, the research combines three complementary approaches (bias reduction, robustness testing, and explainability) rather than relying on any single solution to ensure trustworthy performance.

Abstract

Acoustic sonar image analysis plays a critical role in object detection and classification, with applications in both civilian and defense domains. Despite the availability of real and synthetic datasets, existing AI models that achieve high accuracy often over-rely on seafloor features, leading to poor generalization. To mitigate this issue, we propose a novel framework that integrates two key modules: (i) a Targeted Contrastive Unlearning (TCU) module, which extends the traditional triplet loss to reduce seafloor-induced background bias and improve generalization, and (ii) the Unlearn to Explain Sonar Framework (UESF), which provides visual insights into what the model has deliberately forgotten while adapting the LIME explainer to generate more faithful and localized attributions for unlearning evaluation. Extensive experiments across both real and synthetic sonar datasets validate our approach, demonstrating significant improvements in unlearning effectiveness, model robustness, and interpretability.

Accepted to CVIP 2025

Key Takeaways

AI models can develop problematic "shortcuts" that undermine real-world performance: The research demonstrates that sonar image classifiers achieved high accuracy in lab settings but relied too heavily on seafloor features rather than learning to identify objects properly, illustrating how accuracy metrics alone don't guarantee reliable AI systems.
"Unlearning" is a practical approach to fixing AI bias after deployment: Rather than retraining models from scratch, the researchers developed a method to selectively remove learned biases from existing models, offering a more efficient way to correct AI systems that have learned to depend on irrelevant background features.
Explainability tools are essential for verifying that AI bias removal actually worked: The framework includes visualization methods that show what the model has "forgotten," allowing researchers and stakeholders to confirm that bias mitigation efforts were successful—not just assume improvements based on performance scores.
Responsible AI in high-stakes domains requires multiple safeguards: For applications like sonar detection in defense and civilian contexts, the research combines three complementary approaches (bias reduction, robustness testing, and explainability) rather than relying on any single solution to ensure trustworthy performance.

Abstract

Accepted to CVIP 2025

Supervised Contrastive Machine Unlearning of Background Bias in Sonar Image Classification with Fine-Grained Explainable AI

Key Takeaways

Abstract

Paper Info

Responsible AI Topics

Links

Supervised Contrastive Machine Unlearning of Background Bias in Sonar Image Classification with Fine-Grained Explainable AI

Key Takeaways

Abstract

Paper Info

Responsible AI Topics

Links