Metarep: improving meta-learning accuracy by learning episodic representation

Abstract Meta-learning algorithms face a fundamental challenge: while they train on episodes from one distribution and test on episodes from different distributions, current approaches do not capture unlabeled representations in episodes. We present MetaRep, a novel meta-representation learning framework that learns unsupervised latent representations within episodes to address this gap. Our approach employs weakly augmented samples for the support set and strongly augmented variants for the query set, utilizing a temperature-scaled cross-entropy loss to prevent overfitting during representation learning. The learned parameters are then fine-tuned through supervised meta-learning, making MetaRep model-agnostic and capable of improving any meta-learning architecture. Through extensive experiments on standard few-shot learning benchmarks, we demonstrate that MetaRep significantly improves accuracy across multiple architectures including MAML (+3.8%) and Relation Networks (+4%). Our method shows particular strength in low-data regimes, achieving state-of-the-art performance with up to 75% fewer labeled samples. Additionally, we empirically validate MetaRep’s effectiveness through ablation studies and transfer learning experiments. Code is available at https://github.com/atik666/representationTransfer .