Quantum Machine Learning
Quantum machine learning puts data into qubit states. The hope is that a quantum model can find patterns that normal models can't. No full, start-to-finish advantage on normal (classical) data has ever been demonstrated. The main reason is the cost of loading the data. This chapter works through the three ways to encode data, with real arithmetic. Then it runs an angle-encoding circuit you can check by hand.
What does quantum machine learning actually propose?
Machine learning fits a flexible formula to data, so it can predict well on new inputs. Think of a spam filter that learns from emails you mark as spam. The quantum version, QML, swaps part of that pipeline for a quantum circuit:
- Encode an input into the state of qubits. A qubit's state has weights called amplitudes. Squared, they give the chance of each result.
- Apply a trainable series of gates. These are the basic steps of a quantum program.
- Measure.
- Hand the statistics to a normal optimizer, which adjusts the gates.
Run the chapter 1 filter before getting excited. The claimed speedup is heuristic for nearly all QML on classical data. (Question 1 finds no proof.) And data loading, question 2, is not a small detail here. It is the central problem, as the arithmetic below shows.
There is one truly different case: quantum data. These are states made by a quantum sensor or another quantum computer. They need no loading step. That case is promising for science and far off for business. Everything in this chapter is about the common claim: an advantage on ordinary classical datasets. See the machine learning reality page for tracked claims.
- How does classical data get into a quantum state?
- Worked example: how does angle encoding work by hand?
- What do two angle-encoded features look like?INTERACTIVE
- What happens if a feature is pushed to its maximum?INTERACTIVE
- What is the kernel idea, in plain terms?
- What is the honest evidence table for QML?
- Where does QML stand on real hardware today?
You’ve read the opening of chapter 4. Pro unlocks the other 7 sections — plus every chapter of every course, with circuits you can run right on the page. That’s $11.99 a month, about the price of a coffee, or $99.99 a year (save 30%). The first chapter of every course, and the whole math course, stay free.