Adaptive Quantum Head Reaches 80.9% Accuracy on Fashion-MNIST

Researchers of the New York University Abu Dhabi have demonstrated that a quantum machine learning system, Adaptive KetGPT-QTL, achieved 80.9% accuracy on the Fashion-MNIST dataset, surpassing the performance of a comparable classical system. The work introduces QSTAR: Quantum Selective Transfer with Adaptive Routing, a new framework that strategically employs quantum processing only when a classical model is uncertain. This approach separates the evaluation of quantum circuit design from assessing the actual utility of quantum computation in machine learning. Specifically, KetGPT #180 improved accuracy by 6.82, 4.31, and 3.03 percentage points on low-confidence samples compared to a classical model, suggesting a key role for quantum heads in handling ambiguous data. A compact KetGPT #160 further reached 81.9% accuracy utilizing only ten quantum parameters and nine gates. QSTAR Framework: Quantum Selective Transfer with Adaptive Routing Architecture-searched quantum circuits demonstrate the greatest utility when deployed as targeted fallback branches for handling uncertain data, rather than wholesale replacements for classical classifiers.

Researchers Saim Rehman, Nouhaila Innan, and Muhammad Shafique have moved beyond simply assessing whether a quantum circuit can function as a classifier, to pinpointing when it offers a genuine advantage.

The team’s methodology involved a resource-aware evaluation, considering not only accuracy but also the number of qubits. QSTAR employs a frozen ResNet18 backbone, a classical convolutional neural network pretrained on ImageNet, to initially process images. A lightweight classical branch then assesses the confidence of the network’s prediction; high-confidence predictions are maintained, while low-confidence samples are directed to a quantum fallback head for re-evaluation. This separation of concerns allows for a more nuanced evaluation of quantum utility, isolating the contribution of the quantum circuit from factors like the backbone network or optimization procedures. Experiments utilizing a frozen ResNet18 backbone on Fashion-MNIST revealed that standard QTL heads reach at most 57.0% accuracy, while the strongest KetGPT head in the main filtered sweep reaches 78.5% accuracy and a 0.785 F1-score. Although the strongest fixed classical head remains higher at 81.6%, selective routing gives the quantum branch a clearer role. At the full-system level, Adaptive KetGPT-QTL reaches 80.9% accuracy and 0.807 F1-score, outperforming the adaptive classical baseline. Further investigation revealed that highly compact quantum circuits can also deliver competitive performance. KetGPT #160, a quantum head utilizing only 10 quantum parameters and 9 gates, reached 81.9% accuracy. This challenges the assumption that achieving quantum advantage necessitates large, complex circuits. Recent advances in quantum machine learning increasingly focus on transfer learning, a technique leveraging classical pre-training to reduce the demands on nascent quantum hardware. A fundamental question remains: under what conditions does integrating a quantum circuit actually improve performance, and can its utility be clearly demonstrated?

Researchers Saim Rehman, Nouhaila Innan, and Muhammad Shafique are now exploring selective routing strategies that deploy quantum inference only when classical models falter, rather than simply substituting classical components with quantum ones. This established a consistent baseline, allowing for focused comparison of different classifier heads. They contrasted manually designed quantum heads with those generated by KetGPT, an architecture-search algorithm, and benchmarked these against parameter-matched classical multilayer perceptrons (MLPs). Crucially, the researchers maintained a common data split and optimization schedule to ensure fair evaluation. Saim Rehman, Nouhaila Innan, and Muhammad Shafique are dissecting the practical utility of quantum machine learning components, moving beyond broad performance claims to pinpoint when a quantum processing unit truly enhances classification accuracy. The core question driving this research isn’t simply whether quantum heads can outperform classical ones, but when they offer a demonstrable advantage, particularly given the limitations of current quantum hardware.

The team’s methodology involved a resource-aware evaluation, considering not only accuracy but also the number of qubits. Standard QTL heads reach at most 57.0% accuracy, a result significantly surpassed by the strongest KetGPT head, reaching 78.5% accuracy and 0.785 F1-score in the main filtered sweep. Although the strongest fixed classical head remains higher at 81.6%, selective routing gives the quantum branch a clearer role. KetGPT Head Performance on Low-Confidence Sample Accuracy The pursuit of quantum machine learning algorithms capable of outperforming classical counterparts has focused heavily on transfer learning approaches, yet a critical question remained unanswered: when is quantum processing truly beneficial? This isn’t about wholesale replacement of classical models, but rather a targeted fallback for the most challenging data points. Saim Rehman, Nouhaila Innan, and Muhammad Shafique evaluated both manually designed quantum heads and those generated using KetGPT, an architecture-search technique, within a frozen ResNet18 transfer-learning setup. Standard QTL heads reach at most 57.0% accuracy, while the strongest KetGPT head reached 78.5% accuracy. Although the strongest fixed classical head remains higher at 81.6% accuracy, selective routing gives the quantum branch a clearer role. Further analysis revealed a surprising efficiency; the team demonstrated that a highly compact circuit could deliver competitive results, suggesting a pathway towards practical quantum machine learning on near-term devices. By freezing the ResNet18 backbone and focusing on the performance of the classifier head, the researchers isolated the contribution of the quantum circuit itself.

The team’s methodology involved a resource-aware evaluation, considering not only accuracy but also the number of qubits. This holistic approach provides a more realistic assessment of the potential for practical implementation. Compact-Circuit Ablation: KetGPT #160 Resource Efficiency The pursuit of quantum advantage in machine learning often assumes increasingly complex quantum circuits are necessary, a notion challenged by recent work demonstrating surprisingly strong performance from remarkably compact designs. While many quantum transfer learning (QTL) studies focus on maximizing circuit depth and qubit count, researchers are now questioning whether substantial quantum resources are always essential to achieve meaningful gains. This focus on circuit compactness stems from a broader effort to disentangle circuit design from actual quantum utility. Using a frozen ResNet18 backbone, they systematically compared manually designed QTL heads with those generated through architecture search using KetGPT. The results revealed that KetGPT #160 reached 81.9% accuracy, a significant achievement given its limited size. Further analysis demonstrated the effectiveness of this approach in handling low-confidence samples, a key area where classical models often struggle. The methodology involved filtering KetGPT circuits based on resource constraints, retaining those with a minimum of 8 qubits and applying limits on trainable parameters and gate count. Each accepted candidate was then integrated into the ResNet18 pipeline, with its trainable parameters and classical readout optimized. This rigorous process allowed for a direct comparison of different circuit topologies under consistent conditions. The identification of KetGPT #160 as a strong fixed-head candidate underscores the potential for developing highly efficient quantum machine learning models that can operate within the constraints of near-term quantum hardware. 👉 More information 🗞 QSTAR: Quantum Selective Transfer with Adaptive Routing ✍️ Saim Rehman, Nouhaila Innan and Muhammad Shafique 🧠 ArXiv: https://arxiv.org/abs/2607.21411 Stay currentSee today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals. Tags: