LeanTTA preprint reports lower errors with memory-efficient adaptation on edge devices

Researchers report that LeanTTA, a framework for adapting machine learning models on resource-constrained devices, achieved a 15.7% error reduction against state-of-the-art test-time adaptation methods. The preprint reports peak memory usage of 11.2MB for ResNet18 and adaptation within an order-of-magnitude of normal inference speeds on-device. LeanTTA updates normalization statistics without backpropagation, avoiding reliance on large batches and historical data. It also combines partial adaptation with quantized module fusion to reduce computation. The reported results span sensor modalities; adaptation still takes longer than ordinary inference.