Learning from a few examples remains a key challenge in machine learning. Despite recent advances in important domains such as vision and language, the standard supervised deep learning paradigm does not offer a satisfactory solution for learning new concepts rapidly from little data. In this work, we employ ideas from metric learning based on deep neural features and from recent advances that augment neural networks with external memories. Our framework learns a network that maps a small labelled support set and an unlabelled example to its label, obviating the need for fine-tuning to adapt to new class types. We then define one-shot learning problems on vision (using Omniglot, ImageNet) and language tasks. Our algorithm improves one-shot accuracy on ImageNet from 87.6% to 93.2% and from 88.0% to 93.8% on Omniglot compared to competing approaches. We also demonstrate the usefulness of the same model on language modeling by introducing a one-shot task on the Penn Treebank.
核心贡献 · Key contributions
提出匹配网络，一种结合度量学习与外部记忆的神经架构，用于单样本学习。 Proposes Matching Networks, a neural architecture combining metric learning with external memory for one-shot learning.
引入一种训练流程，通过从标签集采样情节来匹配测试条件，实现元学习。 Introduces a training procedure that matches test conditions by sampling episodes from label sets, enabling meta-learning.
在 Omniglot（93.8%）和 ImageNet（93.2%）上达到最先进的单样本准确率，超越基线。 Achieves state-of-the-art one-shot accuracy on Omniglot (93.8%) and ImageNet (93.2%), outperforming baselines.
在 ImageNet、miniImageNet 和 Penn Treebank 的语言建模任务上定义了新的单样本基准。 Defines new one-shot benchmarks on ImageNet, miniImageNet, and a language modeling task on Penn Treebank.
证明非参数注意力机制允许快速适应新类别而无需微调。 Demonstrates that non-parametric attention mechanisms allow rapid adaptation to new classes without fine-tuning.
展示全上下文嵌入（FCE）通过编码支持集上下文来提升性能。 Shows that Full Contextual Embeddings (FCE) improve performance by encoding support set context.
局限 · Limitations
计算成本随支持集大小线性增长，限制了大规模集的扩展性。 Computational cost grows linearly with support set size, limiting scalability to large sets.
在细粒度分类任务（如 ImageNet 狗子集）上性能下降，归因于标签偏差。 Performance degrades on fine-grained classification tasks like ImageNet dogs subset due to label bias.
模型假设支持集类别与训练类别不相交；若分布偏离则失效。 Model assumes support set classes are disjoint from training classes; fails if distributions diverge.
语言建模结果（32-38%准确率）远低于 Oracle LSTM，表明 NLP 中单样本能力有限。 Language modeling results (32-38% accuracy) are far below oracle LSTM, indicating limited one-shot capability in NLP.
训练时需要仔细的情节采样以匹配测试分布，限制了泛化能力。 Requires careful episode sampling during training to match test distribution, limiting generalization.