Supervised Learning on English AI Terms Dictionary

Labeled data

Sat, 18 Jul 2026 10:04:23 +0000

Definition

Labeled data consists of input samples paired with corresponding ground truth labels, serving as the foundation for supervised machine learning. It allows algorithms to learn the mapping between inputs and outputs by minimizing prediction errors during training. High-quality labeled data is critical for model accuracy, but its creation often requires significant human effort and domain expertise to ensure correctness and consistency across the dataset.

Summary

Data where the correct output or target value is provided alongside the input features.

Empirical risk minimization

Sat, 18 Jul 2026 09:56:53 +0000

Definition

Empirical Risk Minimization (ERM) is the standard objective function for training supervised learning models. It involves selecting a hypothesis from a class of functions that minimizes the average error (loss) calculated on the available training dataset. While ERM aims to fit the data well, it must be balanced with regularization techniques to prevent overfitting, ensuring that the model generalizes effectively to unseen data rather than merely memorizing noise in the training set.

Data annotation

Sat, 18 Jul 2026 09:52:41 +0000

Definition

This critical step involves attaching meaningful metadata to raw data points so that algorithms can learn the relationship between input and output. For example, bounding boxes around objects in images or sentiment labels for text reviews. High-quality annotation is essential for the performance of supervised learning models, as the model’s ability to generalize depends directly on the accuracy and consistency of these labels.

Summary

Data annotation is the process of labeling raw data, such as images or text, to make it suitable for supervised machine learning training.

Binary classification

Sat, 18 Jul 2026 09:48:19 +0000

Definition

Binary classification is a fundamental machine learning problem where the output variable is categorical with exactly two possible outcomes, such as true/false or spam/not spam. Algorithms like logistic regression, support vector machines, and decision trees are commonly used. The model learns a decision boundary that separates the two classes based on training data features, enabling predictions for new, unseen instances.

Summary

A supervised learning task where the goal is to predict one of two possible classes for each input instance.

Active learning

Sat, 18 Jul 2026 09:44:54 +0000

Definition

Active learning reduces the amount of labeled data required by allowing the model to choose the most informative instances for human labeling. Instead of passively receiving random samples, the algorithm identifies regions of high uncertainty or potential impact and requests labels specifically for those cases. This iterative process significantly lowers annotation costs and accelerates convergence, making it ideal for scenarios where data labeling is expensive, time-consuming, or requires specialized expertise.

Multiple instance learning

Sat, 18 Jul 2026 09:41:54 +0000

Definition

Multiple Instance Learning (MIL) addresses scenarios where data is grouped into ‘bags’ with a single label, while individual instances within those bags remain unlabeled. A bag is typically positive if at least one instance is positive, and negative only if all instances are negative. This technique is crucial when precise labeling of individual data points is costly or impossible, allowing models to learn from coarse-grained supervision signals effectively.

Summary

A weakly supervised learning paradigm where labels are assigned to bags of instances rather than individual instances.