計畫名稱：考慮紋理特徵的高效率複數型類神經網路深度學習架構的研究
執行起迄：2017/08/01~2018/10/31
總核定金額：733,000元
中文摘要：捲積型類神經網路(convolutional neural network，以下簡稱CNN) 以多重深度學習之CNN 類神經網路串接 進行深度訓練，對於高維度的複雜問題，可以減少輸入特徵維度的嘗試錯誤，同時獲得有效的為問題表 示特徵與訓練模型。但是CNN 中的捲積計算量大，造成後續訓練時間長；捲積計算僅凸顯訊號最大值， 並不等於訊號重要性，當作為辨識特徵時，往往會遺漏與辨識相關但非最大強度的特殊訊號，反而降低 了辨識效果。再者，CNN 中使用傳統的倒傳遞類神經網路，但因倒傳遞訓練耗時，精確度不穩定，在如 此多層次的訓練架構中，進行深度學習時須耗費巨大的訓練時間。本研究希望改善因為捲積計算的限制 所造成的特徵擷取受限問題，並改善深度類神經網路架構的計算效率問題。我們所提出的方法是以紋理 (texture) 計算取代卷積計算，做為計算特徵；並以複數值類神經網路(Complex-Valued Neural Network) 取 代效率與精確度較差的傳統ANN；最後計算架構中設計符合計算方法的多核心CPU 與GPU 混合型深度 學習加速機制。針對上述議題，本計劃為系列研究的第一年，以「紋理特徵自動擷取與選擇的深度學習 神經網路」為題，主要工作是設計一種自動擷取與選擇資料集中的特徵紋理，取代卷積計算所擷取的單 一的特徵。預計設計以權重式類神經網路架構為不同紋理特徵自動擷取與選擇的機制，作為深度學習的 前端處理。
英文摘要：Deep learning is an emerging technique of machine learning, which has been applied for pattern or image recognition successfully. Convolutional neural network (CNN) is the core technique of deep learning. CNN performs convolution on training data for extracting most significant elements as the features of the target patterns. Then, supervised learning is performed on multi-dimensional back-propagation neural networks. Learning with CNN outperforms traditional learning methods; however, there are some deficiencies. Feature elements extracted by convolution are significantly strong, but may not be important, signals for recognition. Furthermore, the back-propagation training process used in CNN is time-consuming and inaccurate. This study intends to improve the feature learning problem and training efficiency and proposes a three-year project. This report presents the results of the first-year project. The operation of convolution is replaced by texture-based computing. A texture extractor will be developed for computing signification and important texture from training data and served as the front-end preprocessor for deep learning. Some experimental results are presented and discussed.