Quantum Machine Learning With Qiskit

If you want to write quantum machine learning code after this book, you need a realistic starting point, not marketing.

What QML usually looks like in practice

In Qiskit, beginner-friendly QML work usually falls into two families:

• quantum kernels
• variational quantum models

The relevant package is qiskit-machine-learning.

Install the extra package

If you want to run the package-specific examples in this chapter:

uv add qiskit-machine-learning

The circuit design ideas, however, are mostly plain Qiskit ideas you already know.

Quantum kernels

The workflow is usually:

1. encode classical data with a feature map
2. compare encoded states through a kernel
3. feed the kernel to a classical learner such as an SVM

This is conceptually close to the circuit skills you already built:

• data encoding is state preparation
• similarity is controlled by phase and basis structure

A tiny feature-map example

from qiskit import QuantumCircuit
from qiskit.circuit import ParameterVector

x = ParameterVector("x", 2)

feature_map = QuantumCircuit(2)
feature_map.h([0, 1])
feature_map.rz(x[0], 0)
feature_map.rz(x[1], 1)
feature_map.cx(0, 1)
feature_map.rz((x[0] + x[1]), 1)
feature_map.cx(0, 1)

print(feature_map)

Do not worry yet about whether this is the best feature map. The point is to see what QML circuits look like: parameterized state-preparation blocks.

Variational models

A variational model usually has:

• a feature map
• a parameterized ansatz
• an optimizer
• a loss function

The circuit layer is something like this:

from qiskit import QuantumCircuit
from qiskit.circuit import ParameterVector

theta = ParameterVector("theta", 4)

ansatz = QuantumCircuit(2)
ansatz.ry(theta[0], 0)
ansatz.ry(theta[1], 1)
ansatz.cx(0, 1)
ansatz.rz(theta[2], 0)
ansatz.ry(theta[3], 1)

print(ansatz)

This is where parameterized circuits and clean Qiskit workflow really matter.

How the earlier chapters support QML

Kernel methods lean on:

• state preparation
• basis changes
• phase reasoning

Variational methods lean on:

• parameterized gates
• entangling layers
• measurement design
• systematic debugging

So QML is not a separate island. It is built from the same circuit vocabulary you have already been learning.

A realistic beginner goal

You do not need to become a QML researcher immediately. A good beginner goal is:

• read a kernel tutorial without getting lost in the circuit layer
• write a small parameterized ansatz
• understand where trainable parameters live
• know how to inspect a circuit before wrapping it in a training loop

If you can do that, the official Qiskit Machine Learning tutorials become much more approachable.

What to watch out for

QML examples can become confusing for reasons that are not specifically “machine learning”:

• feature-map parameters versus trainable parameters
• classical preprocessing versus quantum encoding
• shot noise versus exact simulation
• overcomplicated ansatzes that are hard to debug

Keep the circuit part small until the workflow is stable.

Checkpoint Exercises

1. Install qiskit-machine-learning with uv add qiskit-machine-learning.
2. Write a tiny parameterized feature map on 2 qubits.
3. Write a simple variational ansatz with a few trainable angles.
4. Explain which earlier chapters in this book support kernel methods and which support variational methods.