Labelled data
Think of this as a study guide for a computer. It is a collection of information where every example comes with the correct answer attached, like flashcards that show both the question and the solution so the computer can learn by checking its work.
A dataset where each input instance is paired with its corresponding ground-truth label. This structured information serves as the essential training signal required to optimize model parameters during supervised learning tasks.
A curated corpus of input-output pairs (x, y) where y represents the ground-truth target variable for a given feature vector x. This annotated data constitutes the supervised signal necessary for minimizing the empirical risk of a predictive model during the training phase.
evolution
- 1957 · historyPerceptron Training
Frank Rosenblatt introduced the Perceptron, requiring manually labeled input-output pairs to adjust weights.
- 1986 · historyBackpropagation Popularization
Rumelhart, Hinton, and Williams demonstrated that multi-layer networks could learn effectively from large sets of labeled training data.
- 2009 · historyImageNet Launch
Fei-Fei Li released ImageNet, a massive, human-labeled dataset that catalyzed the modern deep learning revolution.
- 2017 · historyRise of Weak Supervision
The Snorkel project introduced programmatic labeling, shifting the focus from manual annotation to automated label generation.