Meta-Learning for Cancer Phenotype Prediction from Gene Expression Data

Deep learning has become an essential element in various applications of technology over the past decades. Deep neural networks are now reaching performance on par with, or even beyond, human-level on a broad range of tasks. However, there are still several concerns and deficiencies that make these models impractical for some real-world applications. One of the important issues comes from a data-efficiency perspective. Most of the deep learning techniques need a large number of training samples in order to achieve a high performance on a given problem. This procedure is far from human general intelligence. Humans are good at learning from a few number of samples and quickly adapting to new tasks. In this work, we leverage the meta-learning framework in which the model can learn novel tasks by developing prior knowledge over past experiences. For this purpose, we propose a Meta-Dataset that contains 174 genomics and clinical tasks. Furthermore, we suggest a meta-model under the few-shot learning regime that can learn new genomics tasks. Finally, a comparison between the performance of the meta-learner and the performance of other classical baselines is also presented