Metalearning with Extremely Few Samples Per Task

It is possible to metalearn a good linear representation using just k+2 samples per task, even though k+1 samples per task is insufficient. The number of tasks required scales linearly in the feature dimension and exponentially in the representation dimension.