New gradient-free method enables test-time adaptation for frozen AI models

Researchers have developed CASTER, a novel gradient-free method for test-time adaptation (TTA) of frozen models. This approach stores source class statistics in a subspace and estimates an affine transformation from target-batch moments to adapt distributions before classification. CASTER requires no backward pass or optimizer state, outperforming k-NN on identical frozen features in most tested settings while using significantly less state. The method also includes a transportability certificate to identify unreliable adaptation scenarios, particularly on corrupted datasets like ImageNet-C, and can be gated to improve performance. AI

IMPACT Enables more efficient adaptation of AI models in resource-constrained or inference-only environments.

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

Read on arXiv cs.AI →

• batch normalization
• frozen models
• ImageNet-C
• k-nearest neighbors algorithm
• Tent
• test-time adaptation

• paper
• model release

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