量子訓練: 從模型壓縮的觀點重新思考混合式量子古典機器學習

Abstract

本論文探討量子計算在參數高效學習中的應用，將量子電路重新定位為參數生成器，而非直接用於推論的模型。我們提出Quantum-Train (QT)架構，利用量子神經網路將需訓練的參數數量從 O(M) 降至 polylog(M)，同時保持推論完全於古典電腦上執行；並進一步提出Quantum Parameter Adaptation (QPA)，將此原理擴展至大型預訓練模型的微調，有效降低微調過程中的參數成本。理論分析涵蓋近似誤差與泛化能力，實驗則驗證了這些方法在影像分類、語言模型微調、洪水預測、強化學習以及颱風路徑預測等任務上的效能。結果顯示，QT與QPA能在大幅壓縮參數的同時維持接近的性能，提供一條以效率與可部署性為核心的量子實用性途徑。透過將量子生成的參數整合進古典機器學習流程，本研究展現了量子與經典協同運作於可擴展且資源高效人工智慧中的潛力。

This thesis explores quantum approaches to parameter-efficient learning, reframing quantum circuits as parameter generators for classical models rather than direct inference engines. We introduce Quantum-Train (QT), which employs quantum neural networks to reduce the number of trainable parameters from O(M) to polylog(M) while keeping inference fully classical, and Quantum Parameter Adaptation (QPA), which extends this principle to fine-tuning large pre-trained models with dramatic reductions in adaptation cost. Theoretical analyses examine approximation error and generalization, while empirical studies validate the frameworks across image classification, language model fine-tuning, flood forecasting, reinforcement learning, and typhoon trajectory prediction. Results show that QT and QPA achieve substantial compression with minimal performance loss, offering a realistic pathway toward quantum utility rooted in efficiency and deployability. By integrating quantum-generated parameters into classical machine learning pipelines, this work highlights the potential of quantum--classical synergy for scalable and resource-efficient AI.