ZOTTA framework uses gradient-free optimization for test-time adaptation

Researchers have developed ZOTTA, a novel test-time adaptation (TTA) framework that utilizes gradient-free zeroth-order optimization (ZOO) to enhance model robustness under distribution shifts. Unlike traditional methods that rely on computationally expensive backpropagation, ZOTTA operates solely through forward passes, making it compatible with non-differentiable models and suitable for edge devices. The framework incorporates distribution-robust layer selection to reduce optimization dimensionality and spatial feature aggregation alignment to stabilize the ZOO process. Experiments demonstrate ZOTTA's effectiveness, showing it can outperform or match backpropagation-based methods while significantly reducing memory usage and improving accuracy on datasets like ImageNet-C. AI

IMPACT Enables more efficient and robust model adaptation on edge devices by eliminating the need for backpropagation.

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

Read on arXiv cs.CV →

• ImageNet-A
• ImageNet-C
• ImageNet-R
• ImageNet-Sketch
• Ronghao Zhang
• synthetic aperture radar
• Zeroth-order optimization with orthogonal random directions
• ZOTTA

• paper
• infra

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