Generalising via Meta-Examples for Continual Learning in the Wild
This addresses the challenge of adapting deep learning systems to evolving temporal data with limited supervision, which is crucial for real-world AI applications, though it appears incremental in combining existing concepts.
The paper tackles the problem of learning continually from streams of unlabeled and unbalanced data with scarce annotations, presenting FUSION, a strategy that outperforms state-of-the-art methods in few-shot and continual learning benchmarks.
Future deep learning systems call for techniques that can deal with the evolving nature of temporal data and scarcity of annotations when new problems occur. As a step towards this goal, we present FUSION (Few-shot UnSupervIsed cONtinual learning), a learning strategy that enables a neural network to learn quickly and continually on streams of unlabelled data and unbalanced tasks. The objective is to maximise the knowledge extracted from the unlabelled data stream (unsupervised), favor the forward transfer of previously learnt tasks and features (continual) and exploit as much as possible the supervised information when available (few-shot). The core of FUSION is MEML - Meta-Example Meta-Learning - that consolidates a meta-representation through the use of a self-attention mechanism during a single inner loop in the meta-optimisation stage. To further enhance the capability of MEML to generalise from few data, we extend it by creating various augmented surrogate tasks and by optimising over the hardest. An extensive experimental evaluation on public computer vision benchmarks shows that FUSION outperforms existing state-of-the-art solutions both in the few-shot and continual learning experimental settings.