ETH IVIA Lab’s Post

🔍 Picking the right explainable AI method for your computer vision task? Wondering about its evaluation reliability? Check out our latest #neurips2024 D&B publication on LATEC, a (meta-)evaluation benchmark for XAI methods and metrics. #xai 🚀 Through LATEC, we showcase the risk of conflicting metrics causing unreliable rankings and propose a more robust evaluation scheme. We critically evaluated 17 XAI methods across 20 metrics in 7,560 unique setups, including varied architectures & input modalities. 🎯 Curiously, the emerging top-performing method is not examined in any relevant related study. Dive into our findings: Paper: https://lnkd.in/eApfZJvg Benchmark: https://lnkd.in/ePYMWv3h   👥 This work was a Team effort involving Carsten Lüth, Udo Schlegel, Till Bungert, Mennatallah El-Assady, and Paul F. Jaeger.