實際情境下訓練數據對模型性能影響的量化分析及輕量級神經網絡的應用

Abstract

在現代醫療體系中，主要醫院具有豐富的資源，而分院的資源相對較少。因此，主要醫院會負責訓練深度學習模型，而分院則將其收集的數據提供給主要醫院進行訓練。 本論文的研究目標是在此背景下，探討分院需要累積多少數據，再和主要醫院的數據一起進行訓練，期望在每間參與醫院的測試集上顯著提升模型性能。除了使用燒燙傷的數據集，我們也在交通工具的數據集上模擬了加入不同訓練資料量的情景。除此之外，考慮到主要醫院需要持續接收並訓練新數據，我們的研究目標也嘗試使用綠色學習的架構，運用多種輕量化的模型，旨在接近原始燒燙傷分割模型的結果，同時降低訓練模型時所需的資源和成本。

The main hospital has abundant resources in the modern medical system, while the branch hospitals have relatively few resources. Usually, the main hospital will train the deep learning model, and the branches will provide the data it collects to the main hospital for training. In this study, our goal is to analyze how much data the branch hospital needs to accumulate and train with the data of the main hospital in this context to significantly improve the model performance on the test dataset of each participating hospital. Besides the burn dataset, we simulated scenarios with varying training data volumes using the transportation dataset. Furthermore, considering that the main hospital needs to continuously receive and train new data, our research goal also attempts to utilize green learning frameworks and various lightweight models. The aim is to approximate the initial burns segmentation model results while reducing the resources and costs required during further model training.