Target
The correct output a model is trained to predict, serving as the learning signal.
In supervised machine learning, a target—also called a label, ground truth, or response variable—is the known output value associated with each training example. When a model processes an input, it produces a prediction, and the discrepancy between that prediction and the target is quantified by a loss function. The optimization process, whether gradient descent in neural networks or likelihood maximization in probabilistic models, systematically adjusts model parameters to reduce this discrepancy across the training dataset. Targets can take many forms: continuous values in regression tasks (such as predicting house prices), discrete class labels in classification (such as identifying whether an email is spam), probability distributions in generative modeling, or sequences of tokens in language modeling.
The quality and construction of targets profoundly shape what a model learns. Noisy or mislabeled targets introduce systematic errors that degrade generalization, a problem studied under the umbrella of learning with noisy labels. In some settings, targets are not directly observed but must be inferred—weak supervision, for instance, uses heuristic labeling functions to generate approximate targets at scale, trading label precision for coverage. In reinforcement learning, the concept extends further: the target becomes a reward signal or a bootstrapped value estimate, as in temporal difference learning, where the model essentially learns to predict its own future targets.
Beyond training, the notion of a target influences how practitioners frame problems. Choosing the right target variable is a design decision with significant downstream consequences—a model trained to predict click-through rates may optimize for engagement at the expense of user satisfaction if those two quantities diverge. This gap between the proxy target used during training and the true objective of interest is a central challenge in applied ML, motivating research into reward modeling, human feedback alignment, and evaluation methodology. Understanding what a model is actually being trained toward is therefore as important as understanding the architecture or algorithm used to train it.
