Explainable AI Comparison Hub
19 papers - avg viability 4.9
Recent advancements in explainable AI (XAI) are increasingly focused on integrating interpretability with practical applications across various domains. For instance, new methodologies like counterfactual training are being employed to enhance model transparency by generating plausible, actionable explanations during the training phase, which could significantly improve decision-making in high-stakes environments. Concurrently, frameworks such as Mechanistic Concept Bottleneck Models are refining how models learn and represent concepts, ensuring that explanations are grounded in domain-specific knowledge and causal relationships. Additionally, approaches like UNBOX are addressing the challenge of interpreting black-box models without direct access to their internal workings, making it feasible to audit and understand AI systems in real-world scenarios. These developments not only enhance the trustworthiness of AI systems but also aim to mitigate biases and improve user alignment, thereby addressing critical commercial challenges in sectors ranging from e-commerce to healthcare. The field is clearly moving toward more robust, interpretable models that can operate effectively in complex, real-world settings.
Top Papers
- ECSEL: Explainable Classification via Signomial Equation Learning(8.0)
ECSEL provides an efficient, explainable classification tool for exposing biases and supporting counterfactual reasoning in datasets, with applications in e-commerce and fraud detection.
- Reasoning-guided Collaborative Filtering with Language Models for Explainable Recommendation(8.0)
Develop an efficient and scalable explainable recommendation system using reasoning-guided collaborative filtering with language models.
- Rethinking Concept Bottleneck Models: From Pitfalls to Solutions(7.0)
CBM-Suite enhances Concept Bottleneck Models with a non-linearity fix and distillation loss, enabling more accurate and interpretable predictions, ready for integration into existing vision systems.
- Fusion-CAM: Integrating Gradient and Region-Based Class Activation Maps for Robust Visual Explanations(7.0)
Fusion-CAM enhances the interpretability of deep learning models by adaptively fusing gradient and region-based Class Activation Maps, offering a more robust and detailed visual explanation.
- Process-Guided Concept Bottleneck Model(7.0)
Process-Guided Concept Bottleneck Model improves transparency and reduces bias in scientific AI applications with interpretable intermediate outputs.
- Counterfactual Training: Teaching Models Plausible and Actionable Explanations(7.0)
Develop an AI model training method that inherently provides plausible and actionable counterfactual explanations, enhancing both transparency and robustness.
- Evaluating Explainable AI Attribution Methods in Neural Machine Translation via Attention-Guided Knowledge Distillation(7.0)
A novel approach to evaluate explainable AI methods in neural machine translation using attention-guided knowledge distillation.
- UNBOX: Unveiling Black-box visual models with Natural-language(7.0)
UNBOX provides a framework for understanding black-box visual models using natural language, enabling auditing and bias detection without requiring internal access.
- Learning Concept Bottleneck Models from Mechanistic Explanations(7.0)
M-CBM extracts and names concepts learned by black-box models to build interpretable concept bottleneck models, improving performance and explainability.
- Training for Trustworthy Saliency Maps: Adversarial Training Meets Feature-Map Smoothing(7.0)
Improve the trustworthiness of AI explanations with a training-centered approach that combines adversarial training and feature-map smoothing for more stable and reliable saliency maps.