深度學習是實現人工智慧的技術，它開創了許多人工智慧的實際應用。其優勢在於它們比起動態或物理模型需要更少的開發時間，且相對地簡單。 本研究使用人工智慧類神經網路的一種計算技術，Echo State Network( ESN )演算法，是屬於Reservoir Computing( RC )的一種計算技術。它是一種簡單，快速和高效的非線性演算法，本研究使用該方法應用於雨量之預測以及測謊之判斷。 雨量是評估水資源、農業、生態系統及水文的重要依據。而使用深度學習演算法來預測是一項很有發展性的方式。本實驗使用 ESN 架構，並對台南地區曾文觀測站(120.497E,23.219N)及高雄潮位站(座標為120.283E, 22.617N)自 2002 年至 2014 年每小時的氣象資料進行分析。並將其與其他神經網路演算法做比較，結果表明 ESN的計算結果優於 MATLAB 類神經網路 toolbox 所提供的數種計算工具。此研究成果對於水庫發電與土石流防災有幫助。 本研究另一主題則是使用深度學習進行測謊的研究。目前普遍的測謊儀是藉由觀測受試者之皮膚電阻、呼吸波與脈搏波(血壓)等三項主要參數來測量人們的心理變化，而本研究則是將受試者之回答錄音，經過在語音辨識( Speech Recognition )以及語者辨識( Speaker Recognition )中常用到的特徵參數，稱為梅爾倒頻譜係數( Mel-scale Frequency Cepstral Coefficients, MFCC )之處理後再經過 ESN 架構進行演練，結果證明使用深度學習在測謊應用上也可以達到不錯的成效，平均正確率可達65%，最高正確率可高達100%，為測謊提供一個新的可行方式。;Based on artificial intelligence deep learning has been applied on many applications. The advantage lies in its less development time consumption and relatively more simple than the dynamic or physical models. This study uses an easy, fast and efficient non-linear algorithm, Reservoir Computing ( RC ), based on Echo State Network ( ESN ) algorithm. We apply this algorithm for the prediction of rainfall and the deception detection. Precipitation is a useful information for assessing vital water resources, agriculture, ecosystems and hydrology. Data-driven model predictions using deep learning algorithms are a promising way for these purposes. In this study, we used ESN to analyze the meteorological hourly data from 2002 to 2014 at the Tainan Zengwen Observatory (120.497E,23.219N) and the sea level station in Kaohsiung (120.283E, 22.617N). We also compared the prediction and observation by using the ESN algorithm and a commercial neuron network MATLAB toolbox. The results show that the ESN can provide a better performance to predict rainfall. Another subject of this study is the realization of deception detection based on deep learning. Nowadays, polygraph is performed by the observation of the subject′s skin resistance, respiratory wave and pulse wave (blood pressure) to measure people′s psychological changes. In this study, the vocal signal of the answer to the subjects is acquired. After the processing of the Mel-scale frequency coefficient (MFCC), which is commonly used in speech recognition and speaker recognition, the signal is treated by ESN. The result shows that the method applied on deception detection can also provide a good result. The average accuracy of the deception detection is 65%. The highest correctness can be up to 100%.
