[Paper Review] Do Quantum Circuit Born Machines Generalize?

Kaitlin Gili, Mohamed Hibat-Allah+3 more|arXiv (Cornell University)|Jul 27, 2022

Quantum Computing Algorithms and Architecture4 citations

TL;DR

This paper investigates the generalization performance of Quantum Circuit Born Machines (QCBMs) beyond mere memorization of training data, using a novel evaluation framework. It demonstrates that QCBMs can generalize effectively to unseen, valid bitstrings—even with only 30% of the training data—and generate higher-quality samples from reweighted distributions, indicating strong potential for practical quantum advantage in optimization tasks.

ABSTRACT

In recent proposals of quantum circuit models for generative tasks, the discussion about their performance has been limited to their ability to reproduce a known target distribution. For example, expressive model families such as Quantum Circuit Born Machines (QCBMs) have been almost entirely evaluated on their capability to learn a given target distribution with high accuracy. While this aspect may be ideal for some tasks, it limits the scope of a generative model's assessment to its ability to memorize data rather than generalize. As a result, there has been little understanding of a model's generalization performance and the relation between such capability and the resource requirements, e.g., the circuit depth and the amount of training data. In this work, we leverage upon a recently proposed generalization evaluation framework to begin addressing this knowledge gap. We first investigate the QCBM's learning process of a cardinality-constrained distribution and see an increase in generalization performance while increasing the circuit depth. In the 12-qubit example presented here, we observe that with as few as 30% of the valid data in the training set, the QCBM exhibits the best generalization performance toward generating unseen and valid data. Lastly, we assess the QCBM's ability to generalize not only to valid samples, but to high-quality bitstrings distributed according to an adequately re-weighted distribution. We see that the QCBM is able to effectively learn the reweighted dataset and generate unseen samples with higher quality than those in the training set. To the best of our knowledge, this is the first work in the literature that presents the QCBM's generalization performance as an integral evaluation metric for quantum generative models, and demonstrates the QCBM's ability to generalize to high-quality, desired novel samples.

Motivation & Objective

• To address the lack of formal assessment of generalization in quantum generative models, particularly QCBMs.
• To investigate whether QCBMs can learn and generalize to unseen, valid samples beyond simple memorization of training data.
• To evaluate the impact of circuit depth and training data size on generalization performance.
• To explore the model's ability to generalize not only to valid samples but also to high-quality, reweighted distributions.
• To establish generalization as a critical evaluation metric for quantum generative models, shifting focus from memorization to learning capability.

Proposed method

• Adopted a recently proposed generalization evaluation framework to quantify QCBM performance on unseen data.
• Trained QCBMs on cardinality-constrained distributions where only bitstrings with a fixed number of 1s are valid.
• Evaluated generalization using validity-based metrics (fraction of generated samples that are valid) and quality-based metrics (performance on reweighted distributions).
• Varied circuit depth and training set size (from 10% to 100% of valid data) to study their impact on generalization.
• Used negative log-likelihood (NLL) and Kullback-Leibler (KL) divergence as auxiliary metrics to assess training fidelity.
• Conducted experiments on a 12-qubit QCBM to analyze scalability trends and resource dependencies.

Experimental results

Research questions

• RQ1 Can QCBMs generalize to unseen, valid bitstrings when trained on a subset of the solution space?
• RQ2 How does increasing circuit depth affect the generalization performance of QCBMs?
• RQ3 What is the minimum amount of training data required for QCBMs to achieve strong generalization?
• RQ4 Can QCBMs learn and generate high-quality samples from a reweighted distribution that differs from the training data?
• RQ5 How does generalization performance in QCBMs compare to memorization performance, and what does this imply for practical quantum advantage?

Key findings

• QCBMs achieve optimal generalization performance when trained on as little as 30% of the valid data in the 12-qubit cardinality-constrained dataset.
• Increasing circuit depth leads to a measurable improvement in generalization performance, indicating that expressivity enhances learning beyond memorization.
• The QCBM successfully generalizes to high-quality, unseen bitstrings from a reweighted distribution, generating samples with better quality than those in the training set.
• Generalization performance is highly sensitive to the number of valid training samples, suggesting that data efficiency is a key factor in model performance.
• The model demonstrates strong validity-based generalization, indicating it learns underlying structural patterns rather than simply memorizing training examples.
• This work is the first in the literature to formally assess QCBM generalization as a core evaluation metric, highlighting its potential for constrained optimization tasks.