@unpublished{Wang:2020, author = {Wang, R and Demiris, Y and Ciliberto, C}, publisher = {arXiv}, title = {Structured prediction for conditional meta-learning}, url = {http://arxiv.org/abs/2002.08799v2}, year = {2020} }
TY - UNPB AB - The goal of optimization-based meta-learning is to find a singleinitialization shared across a distribution of tasks to speed up the process oflearning new tasks. Conditional meta-learning seeks task-specificinitialization to better capture complex task distributions and improveperformance. However, many existing conditional methods are difficult togeneralize and lack theoretical guarantees. In this work, we propose a newperspective on conditional meta-learning via structured prediction. We derivetask-adaptive structured meta-learning (TASML), a principled framework thatyields task-specific objective functions by weighing meta-training data ontarget tasks. Our non-parametric approach is model-agnostic and can be combinedwith existing meta-learning methods to achieve conditioning. Empirically, weshow that TASML improves the performance of existing meta-learning models, andoutperforms the state-of-the-art on benchmark datasets. AU - Wang,R AU - Demiris,Y AU - Ciliberto,C PB - arXiv PY - 2020/// TI - Structured prediction for conditional meta-learning UR - http://arxiv.org/abs/2002.08799v2 UR - http://hdl.handle.net/10044/1/83939 ER -