Efficient Test-Time Adaptation through Latent Subspace Coefficients Search
No backprop, no mini-batches—63x faster adaptation on edge devices.
A new paper from researchers including Xinyu Luo and Haoliang Li proposes ELaTTA, a method for adapting machine learning models to distribution shifts at test time without gradients. Traditional test-time adaptation (TTA) methods rely on backpropagation, activation buffering, or test-time mini-batches, making them unsuitable for edge deployment due to high latency and memory. ELaTTA solves this by freezing all model weights and instead optimizing a low-dimensional coefficient vector in a principal latent subspace derived offline via truncated SVD. At inference, it uses CMA-ES (Covariance Matrix Adaptation Evolution Strategy) to maximize prediction confidence, effectively smoothing the objective and stabilizing decisions near boundaries.
ELaTTA achieves state-of-the-art accuracy under both strict and continual single-instance protocols across six benchmarks and multiple architectures. Compared to leading TTA methods, it reduces compute by up to 63x and peak memory by up to 11x. The team also demonstrated on-device deployment on a ZYNQ-7020 platform, proving its viability for real-world edge applications. This gradient-free, single-instance approach makes robust adaptation practical for resource-constrained environments like drones, smartphones, and IoT devices.
- ELaTTA eliminates backpropagation and mini-batches, using only CMA-ES optimization on a low-dimensional latent coefficient vector.
- Achieves up to 63x compute reduction and 11x peak memory reduction vs. prior TTA methods.
- Deployed successfully on a ZYNQ-7020 edge platform, showing real-world feasibility.
Why It Matters
Enables robust model adaptation on resource-constrained edge devices without expensive backpropagation or memory buffers.