发表机构
New York University Abu Dhabi (NYUAD); NYUAD Research Institute（纽约大学阿布扎比分校; 纽约大学阿布扎比分校研究院）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
研究量子转移学习中量子分支何时有用的问题，提出QSTAR框架，通过选择性路由让量子分支作用更清晰，在Fashion-MNIST实验中，该框架下的KetGPT头表现良好，在全系统层面优于自适应经典基线，还确定了更强的固定头候选者。
AI 中文摘要
量子转移学习（QTL）通常通过用固定的变分量子头替换经典分类器来评估，但这隐藏了一个关键问题：量子分支何时真正有用？我们提出了QSTAR：具有自适应路由的量子选择性转移，这是一个选择性QTL框架，它保留高置信度的经典预测，仅将低置信度样本路由到备用分支。在Fashion-MNIST上使用冻结的ResNet18主干，我们在常见的数据分割和优化计划下比较了手动设计的QTL头、KetGPT设计的量子头和参数匹配的经典基线。标准QTL头的准确率最高达到57.0%，而在主要过滤扫描中最强的KetGPT头达到78.5%的准确率和0.785的F1分数。虽然最强的固定经典头仍更高，为81.6%，但选择性路由使量子分支的作用更清晰。在低置信度样本上，KetGPT #180在0.70、0.80和0.90的阈值下比参数匹配的MLP备用提高了6.82、4.31和3.03个百分点的准确率。在全系统层面，自适应KetGPT-QTL达到80.9%的准确率和0.807的F1分数，优于自适应经典基线。单独的紧凑电路消融确定KetGPT #160是更强的固定头候选者，仅用10个量子参数和9个门就达到了81.9%的准确率。这些结果表明，通过架构搜索的量子头作为不确定输入的目标备用分支最有用，而不是经典分类器的统一替代品。
英文摘要
Quantum transfer learning (QTL) is often evaluated by replacing a classical classifier with a fixed variational quantum head, but this hides a key question: when is the quantum branch actually useful? We propose QSTAR: Quantum Selective Transfer with Adaptive Routing, a selective QTL framework that keeps high-confidence classical predictions and routes only low-confidence samples to a fallback branch. Using a frozen ResNet18 backbone on Fashion-MNIST, we compare manually designed QTL heads, KetGPT-designed quantum heads, and parameter-matched classical baselines under a common data split and optimization schedule. Standard QTL heads reach at most 57.0% accuracy, while the strongest KetGPT head in the main filtered sweep reaches 78.5% accuracy and 0.785 F1-score. Although the strongest fixed classical head remains higher at 81.6%, selective routing gives the quantum branch a clearer role. On low-confidence samples, KetGPT #180 improves accuracy over a parameter-matched MLP fallback by 6.82, 4.31, and 3.03 percentage points at thresholds of 0.70, 0.80, and 0.90. At the full-system level, Adaptive KetGPT-QTL reaches 80.9% accuracy and 0.807 F1-score, outperforming the adaptive classical baseline. A separate compact-circuit ablation identifies KetGPT #160 as a stronger fixed-head candidate, reaching 81.9% accuracy with only 10 quantum parameters and 9 gates. These results suggest that architecture-searched quantum heads are most useful as targeted fallback branches for uncertain inputs rather than uniform replacements for classical classifiers.
