Does it require manually labeled examples?

Not necessarily; later observations can supply targets.

Is it the same as reinforcement learning?

No. Learning to predict transitions and learning behavior from rewards are distinct, though they can be combined.

Must the future be uniquely predictable?

No. A model may need to represent uncertainty or multiple plausible outcomes.

AI Glossary / Graphs and symbolic AI / Advanced

Predictive learning

Definition

Predictive learning, in the world-model sense associated with Yann LeCun, learns from observations to predict aspects of an environment's future state. The learning signal can come from what subsequently happens, rather than from separately supplied category labels. A model may predict observations, useful representations or consequences conditional on actions. The term here refers to learning about evolving environments, not merely to every algorithm that outputs a prediction.

What it means

Watching a sequence provides more information than viewing isolated examples. Repeated comparisons can help it learn regularities in how the environment changes. Such knowledge may support anticipation and planning, but learning a predictor does not automatically create a capable planner.

How it works

Construct training examples from sequences, keeping earlier information as input and later information as the prediction target. When actions affect what happens, include action information where appropriate. Train the model to reduce a defined discrepancy between its prediction and the observed target, or between representations of them.
The target need not be every pixel of a future image. A representation can emphasize information useful for later decisions. In LeCun's proposed architecture, predictive world models and learned abstract representations are central research ideas. More generally, uncertain environments may require several possible futures rather than one deterministic forecast. Evaluation should separate prediction on new observations from prediction on training sequences, and should examine repeated forecasts when model outputs become inputs for later steps.

Three practical examples

A robot learns how object positions change after recorded movements. Its model can help compare candidate actions, provided predictions remain reliable for those actions and situations.
A video system learns representations by predicting information from later frames. The task supplies a learning signal from the sequence without requiring a person to label every frame with a category.
An equipment-monitoring system predicts subsequent sensor readings from recent readings and operating settings. Differences between prediction and observation can prompt investigation, but an unexpected value is not, by itself, a diagnosis of a particular fault.

Think of it this way

Imagine learning a game by watching what happens after each move and checking your expectations. Gradually, you form a working picture of its dynamics. That picture can help you imagine possibilities, but it may fail when a new rule or unfamiliar situation appears. Predicting what happens and deciding what should happen remain different problems.

Worked example

Consider a simple recorded system with a numeric state x and an action a. Four observed transitions are: x = 0, a = 1, next state 2; x = 2, a = 1, next state 4; x = 4, a = -1, next state 2; and x = 2, a = -1, next state 0.
Fit the restricted model next state = x + b times a. The quantity to predict is the observed change in state. Here the changes are 2, 2, -2 and -2. Least squares gives b as the sum of action times observed change divided by the sum of squared actions. The numerator is 2 + 2 + 2 + 2 = 8, and the denominator is 4, so b = 2. This model fits all four recorded transitions exactly.
Compare a persistence baseline that predicts the next state will equal the current state. Its errors have magnitude 2 on every transition, giving mean squared error 4. The fitted model's training mean squared error is zero. That comparison establishes fit on these records, not generalization to every possible state.
Starting from x = 3, the fitted model predicts that actions 1 and then -1 produce states 5 and then 3. Those are simulated consequences under the learned rule.
Now suppose a separate evaluation environment follows next state = x + 1.8 times a. Three consecutive actions of 1 from x = 3 produce actual states 4.8, 6.6 and 8.4. An uninterrupted rollout of the fitted model predicts 5, 7 and 9, ending 0.6 above the actual state. If the model receives each actual state before making a fresh one-step prediction, each error is instead 0.2.

Illustrative case study

In a fictional machine-control prototype, the team initially advertises zero prediction error using only the four fitting records above. A reviewer asks for new trajectories and checks whether future values have accidentally entered the inputs.
The team evaluates the changed-dynamics example and reports the one-step and three-step results separately. It also tests action sequences absent from training and compares predictions with a persistence baseline. Before using the predictor to choose actions, it defines the control objective and tests the resulting decisions. The exercise distinguishes learning transitions, evaluating forecasts and selecting behavior, rather than treating a perfect training fit as evidence of complete autonomous competence.

Common misconceptions and mistakes

Predictive learning is not necessarily supervised classification under a different name. In this usage, observations can provide their own targets through time. 'Unsupervised' also does not mean that there is no training objective; self-supervised methods construct targets from available data.
Accurate prediction does not automatically establish causal effects under new interventions. A model trained on recorded actions may fail for unfamiliar action choices. Another mistake is assuming that predicting every observable detail is always necessary. Useful representations can support prediction at a chosen level of abstraction, although discarding decision-relevant information can also harm performance.

Business use

Predictive learning can support research into operational simulation, robotics and anticipatory monitoring. Specify what state information is predicted, the forecast horizon and the actions represented in the data. Compare with simple baselines and evaluate unfamiliar operating conditions. If the model informs a decision, assess that decision separately from average forecast error. Maintain a clear distinction between demonstrated capability in a bounded task and broader claims about an intelligent system's understanding of the world.

Check your understanding

Answer: What fitted coefficient does the example produce? b = 2.

Why is zero training error insufficient?

Answer: It tests only the fitting records and their conditions.

What causes the three-step error of 0.6?

Answer: The uninterrupted rollout repeatedly applies a model whose increment exceeds the evaluation environment's by 0.2.

Related terms

Self-supervised learning
Representation learning
World model
temporal prediction
forward model
model-based planning
rollout
distribution shift

Learn next

Repeat the rollout comparison with alternating actions. Then consider which errors cancel in this specific linear example and why cancellation should not be assumed for other dynamics.

Sources

AIMA glossary, predictive learning: https://github.com/aimacode/aima-glossary/blob/master/glossary.md LeCun, Berkeley colloquium abstract, From Machine Learning to Autonomous Intelligence: https://eecs.berkeley.edu/research/colloquium/220927-2/ Sobal and colleagues, Separating the World and Ego Models for Self-Driving: https://arxiv.org/pdf/2204.07184

Previous: Predictive estimation Next: Predictive ML