Must the energy function be a neural network?

No. The framework can use different kinds of parameterized or specified functions.

Is every energy-based model generative?

No. Energy-based formulations can support discriminative and structured-prediction tasks as well as generative modeling.

Can two systems' raw energy values be compared directly?

Usually not without a justified common scale and interpretation.

AI Glossary / AI and machine learning / Advanced

Energy-based model

Definition

An energy-based model assigns a numerical energy score to a possible configuration or an input-output combination. Lower energy represents greater compatibility under the model, and inference seeks suitable low-energy configurations.

What it means

A model does not always need to begin by assigning a fully normalized probability to every answer. It can instead learn a score that makes compatible combinations preferable to incompatible ones.

The word energy is borrowed from mathematical and physical ideas, but the score need not represent electricity, calories or another directly measured physical quantity. Its meaning is defined by the model.

For example, a system might score how well a proposed label matches an image or how well several proposed output values fit together. The desired result should receive a lower score than unsuitable alternatives under the chosen convention.

An energy value is not automatically a probability or a universal measure of correctness. Its scale, interpretation and usefulness depend on the training procedure and the way candidate answers are searched or compared.

How it works

The designer specifies an energy function, which may contain learned parameters. It evaluates inputs, proposed outputs and sometimes hidden variables that represent unobserved aspects of the configuration.

Training adjusts the function so that desired configurations are favored relative to undesirable ones. Different energy-based methods use different losses and ways of selecting or generating comparisons.

At prediction time, the system searches for a low-energy output or samples configurations according to a model built from the energy. Search itself can be difficult when the space of possibilities is large or contains many local minima.

Some formulations connect energy to probability through an exponential transformation and a normalizing quantity called a partition function. Computing that normalization can be challenging, which is one reason it is important to distinguish an energy score from an already calibrated probability.

Three practical examples

Structured prediction: a model scores combinations of output labels so that their relationships can influence which overall answer is preferred.
Image analysis research: candidate interpretations of an image receive compatibility scores based on learned visual information.
Generative modeling research: an energy function helps define which data configurations are preferred, with a sampling procedure used to explore them.

Think of it this way

Think of a landscape in which plausible solutions sit in valleys and implausible ones sit higher up. Finding a solution means searching for a suitably low point rather than reading a ready-made answer from a list.

The analogy has a limit. The landscape is a learned mathematical construction. A deep valley can reflect poor training or a modeling mistake, and finding the lowest available score does not prove real-world truth.

Worked example

Comparing scores without inventing confidence

Suppose a fictional model assigns energies of 1.2, 2.8 and 4.0 to three candidate labels for a prepared input. Under the lower-is-better convention, the first candidate is preferred among those three.

It would be wrong to report 1.2 as a 1.2% error rate or convert the gap between 1.2 and 2.8 directly into a confidence percentage. The raw values do not support those interpretations.

The team also checks whether an important fourth candidate was omitted from the search. Selecting the best of an inadequate candidate set can still produce an unsuitable result.

The numbers are invented. The example illustrates two separate responsibilities: learning a useful scoring function and searching an appropriate set of alternatives. Evaluation must cover both.

Illustrative case study

A fictional research team develops an energy-based system for assigning labels to connected document fields. The scoring function rewards combinations seen in training, but a fast search routine sometimes stops at a poor local solution.

The team initially tries to retrain the model. A closer investigation separates scoring quality from search quality and finds that different search settings produce better candidates under the same function.

The invented lesson is that a model's objective and the procedure used to find an answer are separate parts of the system. A failure in inference may not be fixed by adding training data alone.

Common misconceptions and mistakes

Energy means power consumption. Here it is a model-defined compatibility score, not necessarily a physical measurement.
Lower energy always means a higher calibrated probability. That interpretation requires a specified probabilistic model and appropriate normalization.
A learned energy function makes search trivial. Finding or sampling good configurations can remain computationally demanding.

Business use

Energy-based models are mainly relevant when discussing specialist machine-learning research or structured prediction systems. Business readers benefit from understanding the separation between scoring alternatives and finding the final answer.

Ask what low energy means, which candidates are considered and how the complete inference process is tested. Compare task outcomes, latency and failure handling rather than treating a favorable internal score as a business result.

Check your understanding

A system chooses the lowest-energy candidate from a list that omits the correct answer. Has minimizing the score guaranteed a correct prediction?

Answer: No. It has selected a preferred candidate within the searched set. The candidate generation, score quality and search procedure all affect the result.

Related terms

Learn next

Read loss surface and optimization. Separate the question of how a model scores an answer from the question of how it searches for one.

Sources

LeCun and collaborators, Energy-Based Models, NYU Computational and Biological Learning Lab: https://cs.nyu.edu/~yann/research/ebm/. Checked 12 October 2026. Numerical scores and document scenarios are original illustrations.

Previous: Endpointing Next: Energy distance