Skip to content

QuantumScope v0.9.4 — One circuit, many classes

Choose a tag to compare

@Onimee58 Onimee58 released this 13 Nov 03:57
· 3 commits to main since this release

QuantumScope v0.9.4 — One circuit, many classes

The Variational Classifier now separates any number of classes, completing multi-class support in the Quantum ML area.

What's new

  • All classes in the Variational Classifier. Choose All classes under Classes. The circuit measures ⌈log₂ k⌉ qubits, and each outcome stands for one class: 2 qubits for three penguin species, 4 for ten digits. Training uses cross-entropy with exact gradients, and BFGS, Adam, validation and early stopping all work as before.
  • Results for many classes: decision regions in each class's colour, a full confusion matrix, per-class ROC curves, balanced accuracy, cross-validation, saved models, and predictions on new files with each prediction's probability.
  • Open in circuit simulator loads the trained circuit with its readout qubits measured and tells you which outcome is which class. Run shots and read the class off the histogram.
  • Clear messages: if there are too few qubits for the number of classes (for example, ten digits on 3 qubits), the app explains why instead of failing.
  • Lessons: the lecture More than two classes has a new section, "One circuit, many outcomes", and its lab has a new step that separates three species with one circuit.

Typical test accuracy (chance in brackets): Iris 92% (33%), Penguins 87% (33%), Digits on 6 qubits 65% (10%), MNIST on 8 qubits about 50% (10%). The Quantum Kernel Classifier scores higher on the harder datasets, because a small circuit has to separate every class at once. This is a useful comparison for class.

Also new since v0.9.0

  • v0.9.1: five new optimizers for parameterized circuits (Adam, SPSA, Rotosolve, quantum natural gradient, simulated annealing), and optimizing with shots, which shows noisy measured energies beside the true energy.
  • v0.9.2: a Quantum annealing experiment. It shows the energy spectrum with the minimum gap, the state during the anneal, the probability of success against the anneal time, and the anneal as a circuit.
  • v0.9.3: the Quantum Kernel Classifier separates any number of classes (one-vs-rest), with four new built-in datasets: Fashion-MNIST, Palmer Penguins, Seeds and Glass Identification.

Please report anything confusing, slow or wrong, with your operating system and a screenshot if you can.

Downloads

  • Linux: .deb
  • Windows: Setup.exe
  • macOS: .dmg

If you already have QuantumScope, use the in-app updater (About → Check for updates).