使用生成式對抗網路及最佳補全序列汲取法之多標籤分類技術

Abstract

本論文的主軸是多標籤分類(Multi-Label Classification)之新技術。隨著機器學習技術的日新月異，基於深層類神經網路(Deep Neural Network)的解決方法陸續被提出，前人的研究指出考慮標籤間的關聯性，是增進模型表現的關鍵。本論文的第一個大方向是以生成對抗網路(Generative Adversarial Network)來模擬標籤關聯性。在此架構下，分類器扮演生成器(Generator)的角色，其輸入是一個物件，輸出是屬於此物件的標籤集(Label set)，鑑別器(Discriminator)則需要學習標籤之間的關聯性，來分辨此標籤集是從生成器產生還是來自真實的資料;分類器不只需要學會標籤和物件間的關係，也需要使產生出的標籤集具有正確的關聯性，以欺騙鑑別器。本論文第二個方向是改進基於遞迴式類神經網路(Recurrent Neural Network)的多標籤分類器;這種模型使用遞迴式類神經網路解碼器來模擬標籤關聯性，並依序預測標籤。然而，此模型在訓練時，需要人為定義的標籤順序，用來將標籤集轉變成標籤序列，為訓練遞迴式類神經網路的目標序列;前人的研究已指出標籤順序對模型表現有相當大的影響，人為強加的順序性也可能會和機器推斷 的標籤關係不一致。因此，本論文提出最佳補全序列汲取法(Optimal Completion Distillation)，使模型不需要標籤順序便可訓練。透過分析實驗數據，我們也證實我們提出的模型不只表現較好，廣泛化能力(Generalization ability)也較強，能夠預測出在訓練集沒有出現過的標籤集。本論文也提供了上述兩種方法在多標籤影像分類、文件分類、環境音分類上相當豐富的測試結果。

Multi-label classification (MLC) assigns multiple labels to each sample. This paper proposes two methods that improves performance of multi-label classifiers. Recent work has shown that exploiting relations between labels improves the performance of multi-label classification. The first direction in this paper is to use Generative Adversarial Network (GAN) to model label dependencies. The discriminator learns to model label dependency by discriminating real and generated label sets. To fool the discriminator, the classifier, or generator, learns to generate label sets with dependencies close to real data. The second direction is to improve state-of-the-art multi-label classifiers , which utilize a recurrent neural network (RNN) decoder to model the label dependency. However, training a RNN decoder requires a predefined order of labels, which is not directly available in the MLC specification. Besides, RNN thus trained tends to overfit the label combinations in the training set and have difficulty generating unseen label sequences. Therefore, we propose a new framework for MLC which does not rely on a predefined label order and thus alleviates exposure bias. We also find the proposed approach has a higher probability of generating label combinations not seen during training than the baseline models. The result shows that the proposed approach has better generalization capability. This paper also provides experimental results on multiple multi-label classification benchmark datasets in different domains, including text classification, image classification and sound-event classification.