Online Few-shot Learning
Online Few-shot Learning (Streaming Meta-Learning from Scarce Labels) · Also known as: online meta-learning, streaming few-shot learning, continual few-shot learning, incremental few-shot learning
Online Few-shot Learning combines the streaming update principle of online learning with the data-efficiency goal of few-shot learning, enabling a model to continuously adapt to new tasks or classes from only a handful of labeled examples as data arrives sequentially — without access to the full historical dataset.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Choose online few-shot learning when labeled data for each new category or task arrives in tiny bursts over time and you cannot afford to retrain from scratch after each update — for example, personalized recommendation as new item types emerge, medical imaging of newly identified subtypes, or robot perception of novel objects encountered in sequence. It is also appropriate when privacy or storage constraints prevent keeping historical training data. Do NOT use it when ample labeled data is available for each class (standard supervised learning will be more accurate and easier to tune), when tasks are entirely static (batch few-shot learning is simpler), or when the episode structure is artificial and the real need is simply continual learning with moderate data volumes.
Strengths & limitations
- Adapts to new classes or tasks from only one to five examples without full retraining.
- Processes data in a streaming fashion, avoiding storage of entire historical datasets.
- Meta-learning across episodes enables increasingly fast adaptation over time.
- Well-suited to privacy-sensitive settings where raw historical data cannot be retained.
- Supports non-stationary distributions: the model can shift its representations as the data stream evolves.
- Requires careful episode construction; if episodes are not representative of the real deployment distribution, meta-training can overfit to artificial task structures.
- Catastrophic forgetting of earlier tasks remains a challenge unless explicit memory replay or regularization is incorporated.
- Computational overhead of the inner-loop adaptation step can be significant for large models in high-throughput streams.
- Evaluation is complex: standard accuracy metrics must be reported per episode type, and aggregation across heterogeneous task distributions is non-trivial.
Frequently asked
How is online few-shot learning different from standard few-shot learning?
Standard few-shot learning assumes a fixed batch of meta-training episodes and trains offline before deployment. Online few-shot learning updates the meta-parameters continuously as each new episode arrives in a stream, so the model improves over time without revisiting all past data.
How is it different from continual learning?
Continual learning typically deals with tasks arriving sequentially with moderate data per task and focuses on preventing forgetting. Online few-shot learning specifically addresses the regime where each incoming task has extremely few labeled examples, making it a stricter data-efficiency constraint layered on top of the sequential update requirement.
What base model architectures work best?
Embedding-based models such as prototypical networks and relation networks are popular because their inner-loop adaptation is lightweight (a single forward pass to compute centroids). Gradient-based methods like MAML can also be used but require careful learning-rate tuning to remain efficient in the online regime.
How should I evaluate an online few-shot model?
Report episode-level accuracy averaged over multiple random seeds and task distributions. Also track performance on held-out episodes from early in the stream to check for forgetting. Use confidence intervals, not just point estimates, given the high variance of few-shot accuracy.
Is catastrophic forgetting always a problem?
It is a risk whenever the meta-parameters are updated without any mechanism to preserve earlier representations. Techniques such as episodic memory replay, elastic weight consolidation, or dedicated memory-augmented modules can substantially reduce forgetting in practice.
Sources
- Finn, C., Rajeswaran, A., Kakade, S., & Levine, S. (2019). Online Meta-Learning. Proceedings of the 36th International Conference on Machine Learning (ICML), PMLR 97, 1920–1930. link ↗
- Javed, K., & White, M. (2019). Meta-Learning Representations for Continual Learning. Advances in Neural Information Processing Systems (NeurIPS), 32. link ↗
How to cite this page
ScholarGate. (2026, June 3). Online Few-shot Learning (Streaming Meta-Learning from Scarce Labels). ScholarGate. https://scholargate.app/en/machine-learning/online-few-shot-learning
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Few-shot LearningMachine learning↔ compare
- Online LearningMachine learning↔ compare
- Semi-supervised LearningMachine learning↔ compare
- Transfer LearningMachine learning↔ compare