IQP Born machines learn from data using circuits up to 484 qubits

Researchers at LMU Munich have trained circuits of up to 484 qubits using a new method for demonstrating a significant scale for these quantum machine learning models. The work reveals that adaptive algorithms building these circuits are fundamentally constrained by their initial “seed,” specifically preventing the addition of gates operating on an even number of qubits.

This constraint arises because a candidate gate only has a nonzero gradient if its generator is the sum of an odd number of linearly independent generators already in the circuit. Using a seed of one-qubit gates plus the edges of the Chow-Liu tree, the team achieved performance matching full circuits on 12 and 16 qubits with only 10 to 13 percent of the parameters.

IQP Born Machines and Data-Driven Learning

A seed incorporating one-qubit gates alongside the edges of a “Chow-Liu tree” circumvents this limitation, allowing the algorithm to evaluate every candidate gate. Expanding the scale to 484 qubits, the team employed a “heredity pool” to manage the search for higher-order gates, maintaining feasibility as circuit size increased. While these circuits reached a lower test loss than a dense two-body circuit, they did so using only 6 percent of the parameters.

Classical Training of High-Qubit IQP Circuits

Training these circuits remains efficient even as qubit counts climb, with classical computers handling the process for systems reaching hundreds of qubits and thousands of parameters while the quantum computer focuses solely on sampling. This division of labor allows for effective training and demonstrates good generalization to new, unseen data for specific tasks. The number of parameters in these circuits, however, can grow rapidly, scaling quadratically with the number of qubits if they include both one- and two-qubit gates.

The structure of these circuits influences how effectively they can be trained; expectation values are calculated using generators within a specific set, contributing to the efficiency of classical training methods. These generators are defined by their interaction with Z-basis measurements, and only those generators that satisfy a particular condition, a non-zero dot product with a circuit state, actively contribute to the calculation.

Gradient Visibility Criterion for Circuit Expansion

The structure of a quantum circuit’s initial “seed” dictates which circuits an adaptive algorithm can ultimately reach, a constraint stemming from the visibility of potential gates during training. This limitation isn’t merely theoretical; exact gradient calculations on 12-qubit circuits confirm the effect. The analysis builds on previous work examining gradients of gates within an existing circuit, extending the understanding to predict the viability of adding new gates.

Expectation values, key to calculating loss gradients, are linked to the Fourier transform of the circuit’s probability distribution, meaning the circuit’s structure fundamentally shapes which gates can be effectively incorporated. The team demonstrated that a candidate gate gains a nonzero gradient only after training has adjusted the circuit’s angles away from their initial values.

Expanding the initial seed can overcome this constraint. “The selection rule and the expansion of the expectation values are known,” the researchers write, “and our contributions are an exact criterion for gates not yet in the circuit together with the trap of affine active sets.” This suggests a pathway toward more efficient quantum machine learning models by strategically designing the initial circuit structure.

Affine Hyperplanes and Active Generator Sets

The structure of an active set dictates its potential for expansion, with circuits remaining “trapped” when a single frequency can identify every active generator. This phenomenon, detailed in the paper’s derivation of the visibility criterion, defines affine active sets where one frequency fully accounts for all gates within the set. Theorem 2 establishes that such sets limit the addition of new gates, effectively confining the circuit’s growth to existing parameters.

A candidate gate’s visibility, its ability to contribute to circuit expansion, depends directly on the number of linearly independent active generators it comprises. The implications extend to seed design, with the paper highlighting a pathway to overcome these constraints.

A seed incorporating one-body gates alongside edges of the Chow-Liu tree, a structure derived from training data, ensures visibility for all candidates from the outset. The analysis reveals that a non-affine spanning active set, one lacking a frequency that identifies all gates, possesses no permanently blind candidates, meaning every potential addition is either immediately visible or becomes visible after incorporating a single suitable gate. The adaptive selection process within the ADAPT-IQP algorithm adds the gates with the largest gradients and is therefore blind to every candidate whose gradient is zero.

Heredity Pool for Scaling to 484 Qubits

Scaling adaptive quantum algorithms to larger qubit counts requires managing the exponential growth of potential gates, and the work demonstrates a strategy for reaching 484 qubits through a carefully controlled “heredity pool”. This pool limits the expansion of candidate gates, prioritizing extensions of existing gates rather than introducing entirely new, complex structures, though higher-order gates on the D-Wave dataset ultimately offer no performance benefit over two-body gates. This approach addresses a critical bottleneck in scaling: the sheer number of potential gates that must be evaluated in each round of the adaptive process.

A pool limited to gates of body order four alone would be too large to efficiently score, necessitating the heredity principle to maintain feasibility. While a seed of one-body gates plus random edges also sees all candidates, it does not outperform the one-body seed on 484 qubits, suggesting the specific structure of the Chow-Liu tree is essential for efficient learning.

The study found that on the Binary Blobs dataset, the one-body seed requires estimation noise to be freed, while on the D-Wave dataset, the heredity pool initially yields only blind two-body gates, necessitating the addition of random edges.

Circuit Performance on the D-Wave Dataset

Analysis of active gates by body order revealed a key difference between the one-body and CL adaptive algorithms. The one-body approach exclusively selected odd body orders, while the CL approach used all orders. The exact Kullback-Leibler (KL) loss was measured on Two Moons to assess performance, with results indicating the effectiveness of the adaptive approach. Further investigation could focus on the specific circuit structures automatically discovered by ADAPT, and explore how higher-order gradients or gradient-free methods might address blind candidates.

The datasets used in the study, Binary Blobs and the D-Wave dataset, originate from the work of Recio-Armengol et al. and Scriva et al., respectively. All code and results supporting the findings are publicly available, allowing for independent verification and further research. “Gradually growing the sparse circuit is more effective than training it from scratch on this dataset,” the researchers state, highlighting the benefits of their adaptive approach to quantum circuit development.

Loss Functions: Kullback-Leibler Divergence and MMD

Scaling the adaptive algorithm, now called ADAPT-IQP, to circuits of 484 qubits requires a heredity pool to manage the computational demands of adding higher-order gates, though analysis reveals these gates offer no performance advantage over two-body gates. The algorithm’s loss gradient relies on expectation values, functions of the circuit’s angles, which are determined by the Fourier transform of the probability distribution being learned. Inversion of this transform allows the algorithm to estimate the underlying distribution from these expectation values and their derivatives.

The choice of loss function is critical. the Kullback-Leibler (KL) divergence directly compares the learned distribution with the true distribution, but requires complete knowledge of the latter. The maximum mean discrepancy (MMD) offers an alternative, enabling training on datasets where the true distribution is only partially known.

Visibility, a property of the expectation values, impacts the loss gradient and can limit the algorithm’s ability to select new gates; a candidate gate’s gradient can be small or zero at certain angles, even if it would otherwise improve the model. The study highlights that the loss gradient of a candidate sums derivatives weighted by the data, and the selection process is constrained by the circuit’s current state.

👉 More information
🗞 Gradient Blindness in Adaptive Architecture Search for IQP Born Machines
✍️ Sebastian Wölckert, Gerhard Stenzel, Michael Poppel, Claudia Linnhoff-Popien and Jonas Stein (LMU Munich, Munich, Germany)
🧠 ArXiv: https://arxiv.org/abs/2611.24174v1

Stay current

See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.

Avatar of The Quant

The Quant

The Quant possesses over two decades of experience in start-up ventures and financial arenas, brings a unique and insightful perspective to the quantum computing sector. This extensive background combines the agility and innovation typical of start-up environments with the rigor and analytical depth required in finance. Such a blend of skills is particularly valuable in understanding and navigating the complex, rapidly evolving landscape of quantum computing and quantum technology marketplaces. The quantum technology marketplace is burgeoning, with immense growth potential. This expansion is not just limited to the technology itself but extends to a wide array of applications in different industries, including finance, healthcare, logistics, and more.

Latest Posts by The Quant: