RADC
A risk-aware dual-cache framework for test-time adaptation in vision-language models. It addresses background bias and unreliable cache admission through a semantic foreground cache and a Gaussian risk admission mechanism.
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.
The semantic foreground cache aggregates category-consistent spatial evidence from CLIP's spatial representations to extract foreground prototypes, thereby mitigating interference from background information.
The foreground cache serves as a complement to, rather than a replacement for, the global cache; the global cache retains overall context while the foreground cache provides purer target semantics.
Gaussian risk admission models multi-view representations as diagonal Gaussian distributions, simultaneously considering class discriminability and feature uncertainty to filter reliable cache candidate samples.
RADC fuses three types of predictions during inference: CLIP zero-shot logits, global cache predictions, and foreground cache predictions.
RADC was evaluated on both cross-domain and out-of-distribution (OOD) benchmarks, achieving consistent state-of-the-art performance.
As a preprint paper, RADC's additional computational overhead, compatibility with different backbone networks, and performance under more extreme distribution shift scenarios remain to be further validated.
The shift from entropy-driven to risk-driven admission is the key distinction between RADC and previous cache-based TTA methods.
All Facts (8)
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%UnverifiedThe semantic foreground cache aggregates category-consistent spatial evidence from CLIP's spatial representations to extract foreground prototypes, thereby mitigating interference from background information.
50%UnverifiedThe foreground cache serves as a complement to, rather than a replacement for, the global cache; the global cache retains overall context while the foreground cache provides purer target semantics.
50%UnverifiedGaussian risk admission models multi-view representations as diagonal Gaussian distributions, simultaneously considering class discriminability and feature uncertainty to filter reliable cache candidate samples.
50%UnverifiedRADC fuses three types of predictions during inference: CLIP zero-shot logits, global cache predictions, and foreground cache predictions.
50%UnverifiedRADC was evaluated on both cross-domain and out-of-distribution (OOD) benchmarks, achieving consistent state-of-the-art performance.
50%UnverifiedAs a preprint paper, RADC's additional computational overhead, compatibility with different backbone networks, and performance under more extreme distribution shift scenarios remain to be further validated.
50%UnverifiedThe shift from entropy-driven to risk-driven admission is the key distinction between RADC and previous cache-based TTA methods.
50%