On the Efficiency of AdapTTA: An Adaptive Test-Time Augmentation Strategy for Reliable Embedded ConvNets

Test-Time Augmentation (TTA) is a popular technique that aims to improve the accuracy of Convolutional Neural Networks (ConvNets) at inference-time. TTA addresses a limitation inherent to any deep learning pipeline, that is, training datasets cover only a tiny portion of the possible inputs. For this reason, when ported to real-life scenarios, ConvNets may suffer from substantial accuracy loss due to unseen input patterns received under unpredictable external conditions that can mislead the model. TTA tackles this problem directly on the field, first running multiple inferences on a set of altered versions of the same input sample and then computing the final outcome through a consensus of the aggregated predictions. TTA has been conceived to run on cloud systems powered with high-performance GPUs, where the altered inputs get processed in parallel with no (or negligible) performance overhead. Unfortunately, when shifted on embedded CPUs, TTA introduces latency penalties that limit its adoption for edge applications. For a more efficient resource usage, we can rely on an adaptive implementation of TTA, AdapTTA, that adjusts the number of inferences dynamically, depending on the input complexity. In this work, we assess the figures of merit of the AdapTTA framework, exploring different configurations of its basic blocks, i.e., the augmentation policy, the predictions aggregation function, and the model confidence score estimator, suitable for the integration with the proposed adaptive system. We conducted an extensive experimental evaluation, considering state-of-the-art ConvNets for image classification, MobileNets and EfficientNets, deployed onto a commercial embedded device, the ARM Cortex-A CPU. The collected results reveal that thanks to optimal design choices, AdapTTA ensures substantial acceleration compared to a static TTA, with up to 2.21x faster processing preserving the same accuracy level. This comprehensive analysis helps designers identify the most efficient AdapTTA configuration for custom inference engines running on the edge.