Abstract
In few-shot learning, typically, the loss function which is applied at testtime is the one we are ultimately interested in minimising, such as themean-squared-error loss for a regression problem. However, given that we havefew samples at test time, we argue that the loss function that we areinterested in minimising is not necessarily the loss function most suitable forcomputing gradients in a few-shot setting. We propose VIABLE, a genericmeta-learning extension that builds on existing meta-gradient-based methods bylearning a differentiable loss function, replacing the pre-defined inner-looploss function in performing task-specific updates. We show that learning a lossfunction capable of leveraging relational information between samples reducesunderfitting, and significantly improves performance and sample efficiency on asimple regression task. Furthermore, we show VIABLE is scalable by evaluatingon the Mini-Imagenet dataset.