標題: 深層對抗式學習於領域調適之研究
Deep Adversarial Learning and Domain Adaptation
作者: 蔡仁傑
簡仁宗
Tsai, Jen-Chieh
Chien, Jen-Tzung
電機工程學系
關鍵字: 深層學習;生成模型;潛在特徵;領域調適;對抗式學習;模式識別;deep learning;generative model;domain adaptation;latent features;adversarial learning;pattern classification
公開日期: 2017
摘要: 近年來在許多不同的理論以及應用上,深度學習透過大量的標記資料的監督式學習發展得相當迅速。然而,收集大量標記資料實際上是相當耗費時間以及資源的。在現實世界中我們並無法總是取得含標記資料,有時候我們只能取得少量的標記資料甚至只有無標記資料。如何利用數據增強(data augmentation)和改進模型的正規化(regularization)是相當關鍵的問題。近年來對抗式學習被廣泛運用於生成近乎真實的資料並且沒有傳統生成模型需要處理馬可夫鍊的收斂問題(mixing problem). 本篇論文主要探討如何利用深層對抗式學習再產生新的訓練資料。 我們的目標在於利用對抗式網路產生新的訓練資料並且應用在產品製造過程的缺陷檢測上。 為了改進系統的效能,我們建立了一個條件式對抗生成網路(conditional generative adversarial networks)並且引入了額外的隱藏代碼(latent code)並且最大化生成樣本與其的互信息(mutual information)。 我們將此非監督式學習模型應用在銅箔缺陷的產品檢測上。 另一方面,轉移式學習也提供了另一種方法來處理標記資料不足的問題而不需透過產生新樣本。轉移式學習根據不同的設定可被細分為各種不同的議題,在本篇論文我們著重的議題為領域調適(domain adaptation)。領域調適主要透過取的源領域(source domain)與標的領域(target domain)資料的共同特徵(shared feature)來將源領域模型調適到標的領域並實現知識的轉移。傳統領域調適方法專注於取得源領域與標的領域的共同特徵而獨有特徵(individual feature)則未被考量。在本篇論中我們提出一個深層混合對抗式學習的架構,此架構可以同時取出源領域與標的領域的共同特徵以及獨有特徵。我們的概念是透過取得含分類類別資訊的共有特徵以及含各領域獨有資訊的獨有特徵來幫助領域調適的學習。我們利用調適網路來取出共有特徵以及分離網路來取出獨有資訊。調適網路和分離網路都是對抗式網路。我們透過在調適網路和分離網路引入混和對抗式學習(hybrid adversarial learning)來分別最佳化分離損失和領域依賴性(domain dependency)上的最小最大化(minimax)問題。調適網路的概念在於我們透過取出無法被一個最佳化的判別器正確分類的特徵作為源領域和標的領域的共同特徵。分離網路的概念在於我們透過取出可以被一個較差的判別器正確的分類的特徵作為獨有特徵。換言之,我們取出來的特徵必須要足夠好來強迫判別器正確分類。 在對抗式生成模型的試驗中,我們比較了在生成銅箔缺陷影像上不同非監督式學習方法的表現。一般而言,缺陷檢測需要相當高的正確率然而產品的缺陷率會式相對較低了,這意味著在鑑別含缺陷以及不含缺陷上,兩者的資料量將相當的不平均,而這將影響分類器的準確性。在領域調適的問題上,我們比較了我們提出的模型在不同的工作上的表現並且展示了對抗式領域分離與調適在文章分類與影像辨識上的效果。
Deep learning has been rapidly developing from different aspects of theories and applications where a large amount of labeled data are available for supervised training. However, in practice, it is time-consuming to collect a large set of labeled data. In real world, we may only observe a limited set of labeled data and unlabeled data. How to perform data augmentation and improve model regularization is a crucial research topic. Recently, adversarial learning has been discovering to generate or synthesize realistic data without the mixing problem in traditional model based on Markov chain. This thesis deals with the generation of new training samples based on deep adversarial learning. Our goal is to carry out the adversarial generation of new samples and apply it for defect classification in manufacturing process. To improve system performance, we introduce the additional latent codes and maximize the mutual information between generative samples and latent codes to build a conditional generative adversarial model. This model is capable of generating a variety of samples under the same class. We evaluate the performance of this unsupervised model by detecting the defect conditions in production process of copper foil images. On the other hand, transfer learning provides an alternative method to handle the problem of insufficient labeled data where data generation is not required. Transfer learning involves several issues owing to different setups. The issue we concern is mainly on domain adaptation. Domain adaptation aims to adapt a model from source domain to target domain through learning the shared representation that allows knowledge transfer across domains. Traditional domain adaptation methods are specialized to learn the shared representation for distribution matching between source domain and target domain where the individual information in both domains is missing. In this thesis, we present a deep hybrid adversarial learning framework which captures the shared information and the individual information simultaneously. Our idea is to estimate the shared feature which is informative for classification and the individual feature which contains the domain specific information. We use adaptation network to extract the shared feature and separation network to extract individual feature. Both adaptation and separation network are seen as an adversarial network. A hybrid adversarial learning is incorporated in the separation network as well as the adaptation network where the according to the minimax optimization over separation loss and domain discrepancy, respectively. The idea in the adaptation network is that we want to extract shared feature that an optimal discriminator cannot tell where feature come from. The idea in the separation network is that we want to extract feature including shared and individual feature which can be separated even by a bad discriminator. In other words the features have to be good enough to force the discriminator to classify them correctly. For the experiment on generative adversarial model, we investigate different unsupervised learning methods for defect detection in presence of copper foil images. In general, defect detection requires very high accuracy but the defect rate usually is relatively low which means the images with and without defect are very unbalanced. We generate the defective images to balance the training data between defective images and non-defective images conditioned on different classes. For the experiments on domain adaptation problem, we evaluate the proposed method on different tasks and show the merit of using the proposed adversarial domain separation and adaptation in the tasks of sentiment classification and image recognition.
URI: http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070450725
http://hdl.handle.net/11536/142404
Appears in Collections:Thesis