Abstract

Real-world data coming from settings like hospital collections for detecting disease experience multiple sources of distributional shifts. These issues affect the performance of diagnostic methods, reducing the quality of service provided and leading to health or economic harm. Deep learning has emerged as a promising method for classification tasks, including diagnostics, and recent progress has led to methods that allow a neural network to adapt network statistics to shifts in specific settings at test time. However, problems arise in these methods adapting to general shifts and domains. In addition, they underperform when data is limited. In our first contribution, we tackle general domain shifts by investigating the key issues leading Test Time Adaptive algorithms to fail under label shift, proposing a means for mitigating these failures. In the second contribution, we tackle few-shot cross-domain adaptation by modifying the affine parameters of the batch norm during few-shot train time, generally enhancing performance. The third contribution parameterizes Scattering Networks, where we enhance a method for low data regimes by providing problem-specific adaptation.