Suppose someone says:
“We are doing quantum machine learning.”
That sentence tells you less than you might think.
Are they taking an ordinary classical dataset and running a quantum classifier?
Are they measuring a quantum device and using a classical neural network to interpret the results?
Are they feeding quantum states directly into a quantum processor?
All three may appear under the same label.
But they are very different scientific problems.
A simple way to organize the landscape is to ask two questions:
- What kind of data do we have?
- What kind of algorithm processes it?
Those choices give four quadrants.
Prefer to watch? Continue with the video below, or keep reading.
The Two Axes
A similar 2×2 taxonomy is widely used in modern QML discussions: classify the problem by whether the data are classical or quantum and whether the processing algorithm is classical or quantum. The broader QML field was consolidated by early surveys and texts and by the influential 2017 Nature review by Jacob Biamonte and collaborators. The exact labels and boundaries vary across authors, so this matrix is best treated as a pedagogical framework rather than a rigid historical standard.
The first axis is the data type.
Classical data include:
- bitstrings,
- vectors,
- images,
- text,
- tables,
- classical measurement records.
Quantum data include:
- quantum states,
- quantum channels,
- coherent outputs of quantum devices,
- quantum information that has not yet been irreversibly reduced to ordinary classical records.
The second axis is the algorithm type.
The learner may be classical.
Or it may use a quantum computer.
Put those two choices together:
| Data | Classical algorithm | Quantum algorithm |
|---|---|---|
| Classical data | CC | CQ |
| Quantum data | QC |
The labels are simple.
The consequences are not.
CC: Classical Data + Classical Algorithms
This is ordinary machine learning.
An image enters a neural network.
A table enters a gradient-boosted model.
Text enters a language model.
Nothing about the data or the computation needs to be quantum.
Why include CC in a quantum machine-learning discussion?
Because it is the baseline.
Any claim that quantum machine learning is useful has to be compared with strong classical alternatives.
A quantum method should not be celebrated for beating a weak classical model.
The relevant question is:
Does the quantum method offer something beyond what the best practical classical strategy can do?
CC keeps us honest.
CQ: Classical Data + Quantum Algorithms
This is the regime most people picture when they hear “quantum machine learning.”
We begin with a classical vector $x$.
Then we encode it into a quantum state:
\[ x \rightarrow |\phi(x)\rangle. \]
A parameterized quantum circuit or other quantum algorithm processes the state.
Measurements produce an output used for classification, regression, generation, or another learning task.
Examples include:
- variational quantum classifiers,
- quantum kernel methods,
- quantum neural networks,
- some quantum generative models.
This is an active and important area.
But it immediately creates a question:
If the data are already classical, how costly is it to put them into the quantum computer?
That is the data-loading problem.
If preparing $|\phi(x)\rangle$ is expensive, an algorithmic speedup inside the quantum processor may disappear once the full pipeline is counted.
We will examine that next.
QC: Quantum Data + Classical Algorithms
Now reverse the situation.
The data source is quantum.
The learning algorithm is classical.
This may sound exotic, but it is very common in quantum experiments.
A quantum device prepares a state $\rho$.
We measure it.
The measurement outcomes become classical records:
\[ \rho \rightarrow \{m_1,m_2,\ldots,m_N\} \rightarrow \text{classical learning}. \]
This includes many tasks in:
- calibration,
- tomography,
- characterization,
- sensing,
- noise analysis,
- quantum-control feedback.
Classical shadows also fit naturally into this broad picture.
The original data source is quantum, but the retained representation used for later prediction is classical.
QC is therefore a major part of AI for Quantum.
The AI may never run on a quantum processor.
It learns from the classical traces left by quantum experiments.
QQ: Quantum Data + Quantum Algorithms
Now we reach the most quantum-native regime.
The input is quantum.
The learner is also quantum.
Instead of measuring the state immediately, we preserve coherent quantum information long enough to process it quantum mechanically:
\[ \rho \rightarrow \text{quantum processor} \rightarrow \text{measurement}. \]
This can allow collective measurements or other operations that are impossible after the state has already been converted into classical records.
Examples may include:
- quantum state discrimination,
- quantum principal-component analysis,
- learning properties of quantum dynamics,
- quantum representation learning,
- direct processing of quantum sensor or communication states.
This regime is especially interesting because there is no classical data-loading problem.
The data are already quantum.
But QQ has its own difficult requirement:
How do we transfer, store, and process the quantum data coherently?
That is a physical-interface problem, not just an algorithm problem.
The Same Word Hides Different Claims of Advantage
The four quadrants also clarify what “quantum advantage” might mean.
In CQ, the question may be:
Can a quantum computer process classical data more effectively than a classical learner?
In QC, the question is different:
How much can we learn about a quantum system using measurements followed by classical processing?
In QQ, we can ask:
Does preserving the quantum information and processing it coherently reduce the computation or the number of experiments needed?
These are not interchangeable claims.
A result in one quadrant does not automatically support a claim in another.
A Simple Example: Learning a Molecule
Imagine we want to predict a molecular property.
CC
A classical molecular descriptor goes into a classical neural network.
CQ
The same classical descriptor is encoded into a quantum circuit, which performs the learning.
QC
A quantum simulation or experiment produces measurement data about the molecule, and a classical model learns from those records.
A quantum state representing the molecular system is processed directly by a quantum learner before final measurement.
All four may be called quantum machine learning in a broad conversation.
But their motivations are different.
Why the Data Axis May Matter More Than the Algorithm Axis
Quantum-computing discussions naturally focus on the algorithm.
Is the circuit quantum?
How many qubits?
How deep?
But for learning problems, the source of the data can matter just as much.
If the data are classical, we must account for the cost of encoding and access.
If the data are quantum, measurement choices determine how much information survives into the classical record.
That is why the horizontal algorithm axis alone is not enough.
The vertical data axis changes the question.
The Quadrants Are Not Equal in Maturity
CC is mature.
CQ is heavily studied, but useful quantum advantage remains difficult to establish broadly.
QC is already important in experimental quantum science because classical analysis of quantum measurement data is unavoidable.
QQ is conceptually powerful but physically demanding.
This asymmetry is useful.
It prevents us from treating “QML” as one technology at one stage of development.
It is better understood as several related research programs.
The Road Ahead
The next three blogs will walk through the quantum-relevant quadrants in order.
First, CQ:
What happens when classical data meet quantum algorithms?
Then QC:
How much can classical AI learn from quantum experiments after measurement?
Then QQ:
What changes if we keep the data quantum before learning?
That sequence reveals an important progression.
At first, quantum computing is used as a processor for ordinary data.
Then AI becomes a tool for understanding quantum experiments.
Finally, the experiment and the learner may become part of one coherent quantum-information pipeline.
Next in the series: CQ: Classical Data on Quantum Computers
0 comments