Just published a paper on few-shot learning for classification. Surprisingly, simpler models beat complex ones with clean data.