Data Efficiency on English AI Terms Dictionary

Semi-supervised learning

Sat, 18 Jul 2026 10:15:05 +0000

Definition

Semi-supervised learning is a hybrid training paradigm that utilizes a small amount of labeled data alongside a large volume of unlabeled data. The core assumption is that the structure of the unlabeled data can help define decision boundaries more effectively than labeled data alone. Techniques such as self-training, co-training, and graph-based methods are commonly used. This approach is valuable when labeling data is expensive or time-consuming, allowing models to achieve performance close to fully supervised methods with significantly fewer labeled examples.

Co-training

Sat, 18 Jul 2026 09:49:31 +0000

Definition

This method leverages multiple distinct feature sets (views) of the same data points. Initially, two classifiers are trained on small labeled datasets from each view. They then predict labels for unlabeled data, selecting high-confidence predictions to augment the training set of the other classifier. This iterative process expands the effective training data size, improving generalization when labeled data is scarce but abundant unlabeled data exists.

Summary

Co-training is a semi-supervised learning algorithm where two views of the data are used to train separate classifiers that iteratively label unlabeled data for each other.

one-shot

Sat, 18 Jul 2026 09:39:14 +0000

Definition

One-shot learning is a specific type of few-shot learning where the algorithm must generalize to new classes or tasks after seeing only one positive example during training. This approach mimics human cognitive abilities, allowing us to recognize objects or concepts after minimal exposure. It relies heavily on feature extraction and similarity metrics rather than extensive statistical pattern recognition over large datasets, making it crucial for scenarios with scarce data.

few-shot

Sat, 18 Jul 2026 09:38:20 +0000

Definition

Few-shot learning enables machine learning models to generalize from very limited data, typically ranging from one to ten examples per class. Unlike traditional supervised learning which requires thousands of samples, few-shot methods leverage pre-trained knowledge or meta-learning strategies to adapt quickly to new tasks. This capability is crucial for real-world applications where collecting large annotated datasets is expensive, time-consuming, or impossible due to privacy constraints.

Summary

A learning paradigm where a model performs a task correctly after being exposed to only a small number of labeled examples.