Résumé: Recent advances in few-shot learning has provided new ways to train deep neural networks with few data points. While specialized models have been proposed for classification, reinforcement learning and generative models, only few studies have addressed few-shot regression. We propose a metric learning algorithm to fulfill that need. The idea is to learn a feature space in which a linear predictor can fit few data points and still be able to generalize very well. We use Kernel Ridge Regression algorithm to find that predictor and to enforce the generalization criterion. The mapping from the input space to the feature space and the trade-off between fitting the data points and generalization are learned through meta-learning. Our experiments constitute successful applications of few-shot learning in the pharmaceutical and clinical settings, and demonstrate that our model is quite competitive with the state of the art.
Note: La présentation sera donnée en français