PALM: Pushing Adaptive Studying Charge Mechanisms for Continuous Take a look at-Time Adaptation
Authors: Sarthak Kumar Maharana, Baoming Zhang, Yunhui Guo
Summary: Actual-world imaginative and prescient fashions in dynamic environments face fast shifts in area distributions, resulting in decreased recognition efficiency. Continuous test-time adaptation (CTTA) immediately adjusts a pre-trained supply discriminative mannequin to those altering domains utilizing check information. A extremely efficient CTTA technique entails making use of layer-wise adaptive studying charges, and selectively adapting pre-trained layers. Nonetheless, it suffers from the poor estimation of area shift and the inaccuracies arising from the pseudo-labels. On this work, we purpose to beat these limitations by figuring out layers by way of the quantification of mannequin prediction uncertainty with out counting on pseudo-labels. We make the most of the magnitude of gradients as a metric, calculated by backpropagating the KL divergence between the softmax output and a uniform distribution, to pick layers for additional adaptation. Subsequently, for the parameters completely belonging to those chosen layers, with the remaining ones frozen, we consider their sensitivity with a purpose to approximate the area shift, adopted by adjusting their studying charges accordingly. Total, this strategy results in a extra sturdy and steady optimization than prior approaches. We conduct intensive picture classification experiments on CIFAR-10C, CIFAR-100C, and ImageNet-C and exhibit the efficacy of our technique in opposition to commonplace benchmarks and prior strategies
