Is it an ordinary neural network running on a quantum computer?

Not necessarily. Quantum learning architectures can differ substantially from classical networks in their operations, information representation and readout.

Can it learn from ordinary spreadsheets?

Some models can use classical numerical features, but those features must be encoded, and the encoding can carry significant cost or discard information.

Does “quantum” guarantee better accuracy or speed?

No. Those are empirical or theoretical claims about specified problems and comparisons, not consequences of the name.

AI Glossary / Deep learning / Advanced

Quantum neural network

Definition

A quantum neural network is a learning model that uses quantum operations in a network-like or trainable computational structure. One common approach uses a parameterized quantum circuit whose measured outputs contribute to predictions and whose parameters are adjusted through an optimization process.

What it means

The term covers several architectures, not a single standard replacement for a conventional neural network. Some models process classical data encoded into quantum states; others learn relationships between quantum states. Hybrid models combine quantum circuit operations with ordinary numerical processing.

A quantum model can also be simulated on a classical computer. That is useful for learning and testing, but running a simulator does not demonstrate a quantum hardware advantage.

The name should therefore prompt three questions: what information enters the model, which part uses quantum operations, and what output is measured? Without those details, “quantum neural network” tells a business reader little about a proposed system's practical capabilities.

How it works

In a common hybrid workflow, classical features are encoded into a prepared quantum state. A circuit applies gates, some controlled by trainable parameters such as rotation angles. Measurements produce outcomes from which the system estimates probabilities or expectation values.

A classical computer computes a loss against the learning objective and proposes parameter updates. Repeated circuit evaluations support optimization; the precise gradient or search method depends on the model.

Architecture, encoding and measurement choices shape what can be learned. Quantum gates are not simply ordinary artificial neurons transplanted into a different processor. Other quantum network proposals use different layer constructions, including transformations between quantum states. Finite measurement samples, device noise and difficult optimization landscapes can all affect the result.

Three practical examples

A teaching classifier encodes a small numerical feature vector into a circuit and trains a measured output to distinguish two classes.
A research team studies a quantum network that approximates a transformation from input quantum states to target states.
An image experiment uses a conventional feature extractor followed by a small trainable quantum component, evaluating the full hybrid model against a classical alternative.

Think of it this way

Think of an adjustable experiment. You prepare an input, set the controls, repeat measurements and adjust the controls based on how the observed results match your objective.

The analogy captures the training loop and the role of repeated measurement. It does not explain quantum mechanics by itself. Quantum states, interference and entanglement obey specific mathematical rules; they are not a license to read every possible answer simultaneously or to bypass the cost of preparing inputs.

Worked example

From measurement counts to a model output

Consider a deliberately simple readout from one measured qubit. Assign outcome zero the value +1 and outcome one the value −1. In an invented batch of 1,000 circuit executions, suppose zero occurs 650 times and one occurs 350 times.

The estimated expectation value is (650 − 350) / 1,000 = 0.30. A classifier might use this number directly, combine it with a bias, or map it to a score. The empirical probability of outcome zero is 0.65.

Neither number automatically means a 65% probability that a business event will occur. That interpretation needs a defined target and evaluation of the complete model. Another batch can produce slightly different counts, even with unchanged settings. This example illustrates finite-sample readout, not a demonstration of quantum speedup or successful training.

Illustrative case study

Fictional Alder Research Studio builds a small hybrid classifier and obtains promising simulator results. Its initial presentation describes the model as faster because the quantum component contains few adjustable parameters.

The evaluation team counts data encoding, repeated measurements, optimization steps and classical preprocessing. It also compares against a tuned classical model using the same features and held-out examples. Device experiments are reported separately from ideal simulation.

The team discovers that its evidence supports a working research prototype, but not its original speed claim. It keeps the experiment and revises the conclusion. The lesson is that a compact circuit diagram and good predictive results do not by themselves establish a useful computational advantage.

Common misconceptions and mistakes

Superposition does not make every prediction available as a readable output at once. Measurement determines what classical information can be extracted.

More qubits or deeper circuits do not automatically make training easier. Some circuit families and objectives suffer barren plateaus, where gradients become extremely small as system size grows. This is a conditional phenomenon, not a statement that every quantum model is untrainable.

A simulator result is not a hardware benchmark. Likewise, a hybrid model's improvement cannot automatically be attributed to its quantum part when preprocessing, parameter budgets or tuning effort differ between comparisons.

Business use

Treat a proposed quantum neural network as a specific experimental system. Request its data representation, circuit or channel architecture, readout, training procedure and full cost accounting.

For a business pilot, choose a measurable question and establish a strong classical baseline before interpreting novelty as value. Include repeatability, noise sensitivity and the number of circuit evaluations. Separate predictive quality, sample requirements, runtime and hardware resources: an improvement in one does not prove improvement in the others. Report the conditions under which any claimed advantage was actually demonstrated.

Check your understanding

A quantum-circuit simulator matches a classical classifier's accuracy on a small dataset. Has the experiment demonstrated quantum advantage?

Answer: No. It demonstrates performance for that simulated model and evaluation. A claim of advantage needs an appropriate comparison and clearly defined resource or performance benefit, including the costs of the complete workflow.

Related terms

Quantum machine learning
qubit
parameterised quantum circuit
variational classifier
quantum measurement
hybrid model
barren plateau
data encoding

Learn next

Learn basic quantum states, gates and measurement first. Then trace a small variational classifier from feature encoding through readout and loss. Compare ideal simulation with finite-shot results, and study how encoding, gradient estimation and classical baselines affect a credible evaluation.

Sources

Sources: Beer and colleagues, Training deep quantum neural networks: https://www.nature.com/articles/s41467-020-14454-2 PennyLane, Variational classifier: https://pennylane.ai/demos/tutorial_variational_classifier McClean and colleagues, Barren plateaus in quantum neural network training landscapes: https://www.nature.com/articles/s41467-018-07090-4 PennyLane, Quantum transfer learning: https://pennylane.ai/demos/tutorial_quantum_transfer_learning/

Previous: Quantum natural language processing Next: Quartile