Multi-Domain Learning by Meta-Learning: Taking Optimal Steps in Multi-Domain Loss Landscapes by Inner-Loop Learning

We consider a model-agnostic solution to the problem of Multi-Domain Learning\n(MDL) for multi-modal applications. Many existing MDL techniques are\nmodel-dependent solutions which explicitly require nontrivial architectural\nchanges to construct domain-specific modules. Thus, properly applying these MDL\ntechniques for new problems with well-established models, e.g. U-Net for\nsemantic segmentation, may demand various low-level implementation efforts. In\nthis paper, given emerging multi-modal data (e.g., various structural\nneuroimaging modalities), we aim to enable MDL purely algorithmically so that\nwidely used neural networks can trivially achieve MDL in a model-independent\nmanner. To this end, we consider a weighted loss function and extend it to an\neffective procedure by employing techniques from the recently active area of\nlearning-to-learn (meta-learning). Specifically, we take inner-loop gradient\nsteps to dynamically estimate posterior distributions over the hyperparameters\nof our loss function. Thus, our method is model-agnostic, requiring no\nadditional model parameters and no network architecture changes; instead, only\na few efficient algorithmic modifications are needed to improve performance in\nMDL. We demonstrate our solution to a fitting problem in medical imaging,\nspecifically, in the automatic segmentation of white matter hyperintensity\n(WMH). We look at two neuroimaging modalities (T1-MR and FLAIR) with\ncomplementary information fitting for our problem.\n