Search papers, labs, and topics across Lattice.
The paper introduces PACE, a backpropagation-free continual test-time adaptation method that optimizes normalization layer affine parameters using a low-dimensional subspace. PACE employs Covariance Matrix Adaptation Evolution Strategy with Fastfood projection to efficiently optimize high-dimensional parameters. The method also incorporates an adaptation stopping criterion and a domain-specialized vector bank to improve runtime efficiency, achieving state-of-the-art accuracy and reducing runtime by over 50% compared to existing backpropagation-free methods across multiple benchmarks.
Backpropagation-free test-time adaptation can be both accurate and efficient: PACE achieves state-of-the-art accuracy while slashing runtime by over 50%.
We introduce PACE, a backpropagation-free continual test-time adaptation system that directly optimizes the affine parameters of normalization layers. Existing derivative-free approaches struggle to balance runtime efficiency with learning capacity, as they either restrict updates to input prompts or require continuous, resource-intensive adaptation regardless of domain stability. To address these limitations, PACE leverages the Covariance Matrix Adaptation Evolution Strategy with the Fastfood projection to optimize high-dimensional affine parameters within a low-dimensional subspace, leading to superior adaptive performance. Furthermore, we enhance the runtime efficiency by incorporating an adaptation stopping criterion and a domain-specialized vector bank to eliminate redundant computation. Our framework achieves state-of-the-art accuracy across multiple benchmarks under continual distribution shifts, reducing runtime by over 50% compared to existing backpropagation-free methods.
