From 49ff42fa632f57336fd3b0f5b151bac23a27ffe9 Mon Sep 17 00:00:00 2001 From: Chris Endemann Date: Fri, 20 Dec 2024 08:38:50 -0600 Subject: [PATCH] Update 7c-OOD-detection-energy.md --- episodes/7c-OOD-detection-energy.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/episodes/7c-OOD-detection-energy.md b/episodes/7c-OOD-detection-energy.md index 0825fba..4a749cc 100644 --- a/episodes/7c-OOD-detection-energy.md +++ b/episodes/7c-OOD-detection-energy.md @@ -61,6 +61,7 @@ Here: The exponential relationship ensures that even small differences in energy values translate to significant changes in probability. If your stakeholders or downstream tasks require interpretable confidence scores, a Gibbs-based probability might make the thresholding process more understandable and adaptable. However, the raw energy scores can be more sensitive to OOD data since they do not compress their values between 0 and 1. + ### 3. Why energy works better than softmax Softmax probabilities are computed as: @@ -527,4 +528,4 @@ The energy-based approach for out-of-distribution (OOD) detection has several st - While energy-based methods excel in many scenarios, challenges include tuning thresholds across diverse OOD classes and ensuring generalizability to unseen distributions. - Transitioning to energy-based detection lays the groundwork for exploring training-time regularization and hybrid approaches. -:::::::::::::::::::::::::::::::::::::::::::::::::: \ No newline at end of file +::::::::::::::::::::::::::::::::::::::::::::::::::