Test-Time Adaptation
A technique that dynamically adapts a model to new environments during inference without retraining. It addresses the distribution shift between training and test data, enabling models to generalize to previously unseen conditions at deployment time.
Timeline (last 90 days)
RADC (Risk-Aware Dual Caching) is a risk-aware dual caching framework for test-time adaptation of vision-language models, consisting of two core components: a semantic foreground cache and Gaussian risk admission.
Cache-based test-time adaptation methods require no backpropagation and have low computational overhead, contrasting with gradient-based TTA methods that require backpropagation, such as TTT and TENT.
Existing cache-based TTA methods mostly rely on entropy to decide whether a sample enters the cache, but entropy cannot reliably reflect sample quality when representations undergo drift.
All Facts (3)
RADC (Risk-Aware Dual Caching) is a risk-aware dual caching framework for test-time adaptation of vision-language models, consisting of two core components: a semantic foreground cache and Gaussian risk admission.
50%UnverifiedCache-based test-time adaptation methods require no backpropagation and have low computational overhead, contrasting with gradient-based TTA methods that require backpropagation, such as TTT and TENT.
50%UnverifiedExisting cache-based TTA methods mostly rely on entropy to decide whether a sample enters the cache, but entropy cannot reliably reflect sample quality when representations undergo drift.
50%