Abstract

In the last few years, several methods have been developed to tackle the few-shot setting that is, learning from a limited amount of data belonging to a specific task. Most approaches have followed a meta-learning paradigm that is, learning-to-learn through exposure to large numbers of training tasks (episodic training) with the aim of generalizing to new unseen tasks at test time. Meta-learning has been accomplished in different ways, via rapid adaption in a few gradient steps (e.g. MAML ), exploiting learned metrics (e.g. ProtoNets), or through probabilistic approaches (e.g. Deep Kernel Transfer, CNA Ps). However, most meta-learning methods are affected by severe issues such as training instability and poor scaling that hinder their performance. In addition, there has been a growing body of empirical research showing that simpler fine-tuning routines are very effective at few-shot image classification, while being much easier to train and deploy. In this talk I will start from these empirical findings, comparing strengths and weaknesses of meta-learners and fine-tuners. I will then introduce technical solutions that could be used as building blocks for a second generation of meta-learners. In particular, I will describe our recent method called LITE (Bronskill et al., NeurIPS 2021) that allows meta-training efficiently on large images, and a new adaptive block called CaSE (Patacchiola et al., NeurIPS 2022) that allows fast adaptation of pretrained models on a context set. I will provide strong empirical evidence showing that methods based on LITE and CaSE are able to achieve state-of-the-art performance on a variety of tasks, including real-world personalization benchmarks such as the recently proposed ORBIT .