New FAMPE method enhances AI model explainability by tuning frequency bands

Researchers have developed a new attribution method called FAMPE (Frequency-Aware Model Parameter Explorer) to improve the explainability of AI models. Unlike previous methods that perturb all frequencies of an input image equally, FAMPE independently adjusts low and high frequencies. This allows the method to adapt to the specific model being analyzed, leading to more faithful attribution maps. Experiments on ImageNet with various convolutional neural network and vision transformer architectures showed that FAMPE significantly improved the Insertion Score compared to the AttEXplore baseline, with further gains achieved by using a probe to automatically select the optimal frequency band balance for each image. AI

IMPACT This new method could lead to more trustworthy and understandable AI models by improving the accuracy of attribution maps.

RANK_REASON The cluster contains an academic paper detailing a new method for AI model explainability. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Ali Yavari
• AttEXplore
• convolutional neural network
• FAMPE
• ImageNet
• Insertion Score
• vision transformer

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →