LeanTTA: A Backpropagation-Free, Stateless Approach to On-Device Test-Time Adaptation

For small, compressed models constrained by edge hardware and limited data storage or availability, test-time adaptation (TTA) can add back robustness against common corruptions that induce covariate shift. However, many TTA methods rely on backpropagation and large batch sizes, which we show fail under constrained data and hardware. In this work, we present LeanTTA, a combination of memory and compute-efficient tactics which allow us to deploy TTA on-device, even under quantized paradigms. LeanTTA consists of two components: (1) stateless, backpropagation-free statistics adjustment using an efficient calculation of the Mahalanobis distance and (2) partial adaptation through layerwise updates. LeanTTA reaches up to a 15.7% relative error reduction with minimal additional compute time and memory required over normal inference, demonstrating the potential of backpropagation-free methods for efficient on-device TTA.