- 學位論文
利用深度學習中的轉移學習來處理錯誤標記的數據
Utilizing mislabeled data by transfer learning in deep neural networks
摘要
深度學習對於設備與資料有著非常高的要求[1]，其中資料更是模型好壞的核心。然而在實際問題上，我們要獲取標記好的資料是不容易的。在我們的論文中就是希望能解決這類型的問題，儘管手上的資料不一定非常的好，但還是能透過一些方法，得到一個結果不差的模型。 我們的實驗主要是在處理無人機影像與衛星影像的地物分析。這些資料數量不多，且不是這麼容易進行標記。我們透過GoogleAPI尋找類似我們實驗的資料集來進行訓練，然而我們收集的資料也不是這麼理想，有20%-30%的比例是標記錯誤的。 為了解決這個問題，我們藉由[2],[3],[4],[5],[6]..等類神經網路來設計我們的模型架構，並參考了[7], [8] ,[9]等非監督式模型。透過[10]可以先假定我們原始資料所訓練的模型是不差的，並結合blockwise的model來輔助我們修改標記，來解決標記錯誤的問題。為了取得更好的效果，我們將這個方法進行疊代，使得這筆GoogleAPI的資料集提升了3%~5%的準確度，並找到了我們想要的模型。 最後，當我們拿到一筆新的未標記資料時，可以在不使用其他資料集的前提下，只需進行一些簡單的標記，就可以利用我們的模型架構和演算法直接訓練出一個適合的模型，依此來解決難以取得有標記資料的問題。
並列摘要
Deep learning has very high requirements for equipment and materials [1], and the data is the core of the model. However, in practical matters, it is not easy for us to obtain the marked information. In our paper, we hope to solve this type of problem. Although the information we have is not good enough, we can still get a model with good results through some methods. Our experiments are mainly in image segmentation of unmanned aerial vehicles and satellite imagery. The amount of this information is less and difficult to label, so we use GoogleAPI to find a data set similar to our experiment for training. However, the data we collected is not so ideal, and 20%-30% is labeled incorrectly. In order to solve this problem, we design our model architecture by [2],[3],[4],[5],[6] and other neural networks, and reference to [7], [8], [9] and other unsupervised models. By [10], we can assume that the model we trained in the original data is not bad, then combine with the block-wise model to assist us in modifying the markup and solve the problem of markup errors. For better results, we use this method to iterate. Then we improved the 3%~5% accuracy in the GoogleAPI dataset and found the model we wanted. Finally, when we get a new unlabeled dataset, we can use our model architecture and algorithms to directly train a suitable model without using other data sets. In order to solve the problem of difficult to obtain labeled data.