人工智慧的研究過去60多年來從未停歇，隨著科技的日新月異，我們希望電腦可以像人類一樣具備學習能力，近年來因電腦圍棋alpha go一戰成名，讓更多人投入機器學習 (Machine Learning) 以及深度學習 (Deep Learning) 之領域，因此也發展出許多不同的網路架構，透過這些網路架構來讓電腦輔助人類對資料進行判斷與分類偵測。 本論文利用深度學習中的膠囊神經網路 (Capsule Neural Network, CapsNets) 作為方法，提出應用於聲音事件偵測的系統。將所提取的特徵，透過向量的方式丟入神經網路進行訓練，除了膠囊網路本身可以有效的辨別重疊事件，我們再將膠囊網路拓展為平行的膠囊網路，使每單個膠囊可以學習到更多的特徵，透過以上方法相比於DCASE 2017的Baseline錯誤率下降約41%，而與DCASE 2017 競賽第一名之架構相比，錯誤率也下降26%左右。 ;The research of artificial intelligence has never stopped for more than 60 years. With the rapid development of technology, we hope that computers can have the same learning ability as human beings. In recent years, more and more people invest in the field of machine learning and deep learning, because of the success of the alpha go. Many different network architectures have been developed to allow computers to assist humans in detecting and classifying data. We used the Capsule Neural Network (CapsNets) in deep learning as a method. Propose a system for sound event detection. The extracted features are sent to the neural network for training through the vector. In addition to capsule network can effectively identify overlapping events, we expand the capsule network into a parallel capsule network, let per capsule can learn more features. Compared with DCASE 2017 Baseline, our proposed method error rate is reduced by about 41%. Compared with the architecture of the first place in DCASE 2017 challenge, the error rate also dropped by about 26%.
