Source-linked AI summary
Quantum semi-supervised generative adversarial network for enhanced data classification
Kouhei Nakaji, Naoki Yamamoto
TL;DR
Quantum machine-learning methods may face data-loading requirements that threaten their potential advantage. This paper proposes qSGAN, pairing a quantum generator with a classical discriminator/classifier, and reports comparable classification accuracy to a deep classical generator in a specific setting, with fewer generator parameters.
Problem
Quantum machine-learning circuits may offer higher expressibility, but data loading or QRAM requirements can threaten that potential advantage.
Source: §1 · p. 1Method
Results
About 67% classification accuracy with the classical neural-network generator is comparable to the four-layer quantum generator, which uses 32 parameters versus 3688.
Source: Comparison with a classical neural… · p. 6Takeaways & Limitations
The connected-disconnected image experiment reports classification accuracy as high as with a deep classical generator and noise tolerance under depolarization noise.
Source: §4 · p. 6The described training setup uses batches with ℓ labeled and m−ℓ unlabeled data.
Source: §2 · p. 2
Abstract
from arXiv · showhide
In this paper, we propose the quantum semi-supervised generative adversarial network (qSGAN). The system is composed of a quantum generator and a classical discriminator/classifier (D/C). The goal is to train both the generator and the D/C, so that the latter may get a high classification accuracy for a given dataset. The generator needs neither any data loading nor to generate a pure quantum state, while it is expected to serve as a stronger adversary than a classical one thanks to its rich expressibility. These advantages are demonstrated in a numerical simulation.
1 Introduction
Quantum machine learning may benefit from quantum circuits’ expressibility, but data-loading requirements can threaten that advantage. The paper introduces qSGAN, which uses a quantum generator to train a classical discriminator/classifier, and outlines numerical tests of its performance.
1 Introduction
- Quantum circuits’ possible higher expressibility motivates quantum machine learning, but data loading or QRAM may erase a potential advantage.The introduction cites quantum classifiers as possible alternatives to conventional classifiers such as support vector machines.Source: §1 · p. 1
- GANs pair a generator of fake data with a discriminator, and adversarial training is studied in quantum machine learning.Prior work includes quantum generator/classifier systems; one approach requires the generator to produce a target pure quantum state [18].Source: §1 · p. 1
- qSGAN pairs a quantum generator with a classical discriminator/classifier whose main goal is to become a good classifier, not to produce a high-quality generator.The quantum system is used to train the classical system; the generator needs neither data loading nor pure-state generation and is expected to be a stronger adversary through rich expressibility.Source: §1 · p. 1
- The numerical study examines circuit size, labeled-data quantity, noise tolerance, and comparison with a classical neural-network generator.The introduction says the generator comparison finds comparable classification performance in a specific condition.Source: §1 · p. 1
2 Algorithm of qSGAN
The qSGAN algorithm uses measurement outcomes from a parameterized quantum circuit as fake samples and trains a classical D/C to judge real versus fake and classify labels. Its batchwise procedure updates the generator and D/C with separate objectives, then returns the trained classifier.
Algorithm overview
- Each batch contains labeled and unlabeled real data; the quantum circuit’s computational-basis measurement produces fake samples with probabilities set by its output state.The generator’s randomness comes from an intrinsic quantum property rather than added random seeds.Source: §2 · p. 2
D/C
- The classical D/C combines real-source likelihood with a c+1-component classifier output for the c classes and fake data.Its shared network function feeds separate final-layer functions for classification and real/fake discrimination.Source: §2 · p. 2
Training rule
- For each batch, real and generated fake data are sent to the D/C, while generator and D/C parameters are updated according to their respective training rules.The generator is trained to have fake data classified as real; the D/C learns real/fake judgments and true labels.Source: §2 · p. 2
Generator update
- The generator objective LG updates circuit parameters θ to make generated samples more likely to be judged real by D.The described gradient estimate uses circuit measurements with a parameter shifted by ±π/2.Source: §2 · p. 2
D/C update
- The D/C objective combines real/fake discrimination and label classification, and its neural-network parameters are updated by gradient descent.The label-classification term uses cross entropy between the predicted class distribution and the true-label distribution.Source: §2 · pp. 2–3
Algorithm output
- After the batchwise training iterations, the trained classifier C(x) is the algorithm’s final output.The procedure updates generator parameters by gradient descent and D/C parameters by back-propagation.Source: §2 · p. 3
3 Numerical demonstration
The numerical simulations test qSGAN on a small image-classification task, varying the number of labeled examples, generator depth, and generator type. They report the strongest quantum-generator accuracy with few labels, comparable accuracy to a classical generator in that setting, and tolerance to depolarizing noise up to a stated level.
Problem setting and result
- The simulation uses 56 binary 1 × 8 images split into eight batches, with two batches for training and six for testing in each trial.Four-fold cross-validation changes the training/test assignment; each batch has ℓ=2 or ℓ=5 labeled examples.Source: Problem setting and result · pp. 3–4
- The quantum generator is an eight-qubit circuit with one or four layers, while the classical D/C uses shared hidden layers and separate classifier and discriminator outputs.The classifier has three output nodes for the two labels and fake data; the discriminator has one output node.Source: Problem setting and result · p. 3, Figure 2 · p. 4, Figure 3 · p. 4
- When only ℓ=2 labeled data are available, the four-layer quantum generator yields the highest classification accuracy; when ℓ=5 of eight image data per batch are labeled, all three generators achieve almost the same accuracy.Source: Problem setting and result · p. 4, Figure 4 · p. 5
Noisy qSGAN
- For ℓ=2, classification accuracy does not fall below the noiseless case when per-layer depolarization is suppressed to p=0.05.The noise experiment applies a depolarizing channel between generator layers and measures final-iteration classification accuracy.Source: Noisy qSGAN · p. 5, Figure 5 · p. 5
Comparison with a classical neural network generator
- For ℓ=2, the classical-generator classifier reaches about 67% accuracy, comparable to the four-layer quantum generator, using 3,688 rather than 32 generator parameters.The paper reports roughly 400 iterations for the classical generator versus about 40 for the quantum generator to reach 67%.Source: Comparison with a classical neural… · p. 6
4 Conclusion
The conclusion presents qSGAN as semi-supervised learning with a quantum generator and classical discriminator/classifier, avoiding data loading and pure-state generation. On connected-disconnected images, it reports classification accuracy comparable to a deep classical generator and noise tolerance under depolarization noise.
4 Conclusion
- qSGAN uses a quantum generator to train a classical discriminator/classifier, without data loading or generating a pure quantum state.The generator is intended to act as a possibly stronger adversary than a classical one.Source: §4 · p. 6
- On the connected-disconnected image dataset, the quantum generator’s rich expressibility contributes to classification accuracy as high as a deep classical neural network generator’s, despite its greater parameter count.Source: §4 · p. 6
- The numerical experiment also exemplifies qSGAN’s noise-tolerant property under depolarization noise.Source: §4 · p. 6