深度神經網路是一廣泛被使用的人工智慧技術，在人工智慧終端裝置應用上，深度神經網路引擎需要符合低耗能、高彈性、及短的端至端等待時間。因此本總計畫『應用於人體姿勢辨識與機器人之可重組深度神經網路引擎』將開發一應用於人體姿勢辨識與機器人之可重組深度神經網路引擎來符合這些要求。所開發之可重組深度神經網路引擎具有以下創新性：1）可支援監督式及強化式學習之深度神經網路；2）可支援不同深度神經網路模型之推論；3）使用混合數位與類比運算單元來降低耗能；4）利用姿勢與行為辨識系統以驗證此引擎之效能。此子計畫『應用於監督式學習之可重組深度神經網路技術』將開發應用於監督式學習深度神經網路之可重組技術。所開發之可重組技術具有以下創新性：1）固定端至端等待時間之可重組技術；2）可支援不同模型之可重組技術；3）可用於混合數位與類比運算單元之可重組技術。 ;Deep neural network (DNN) is one widely used artificial intelligence (AI) technique. The DNN engine in edge devices for the AI applications should meet the requirements of low energy, high flexibility, and short end-to-end latency. Therefore, we attempt to develop a reconfigurable DNN (RDNN) engine for the applications of human posture recognition and robotics to meet the requirements under the grand project entitled “Reconfigurable Deep Neural Network Engine for human posture recognition and robotics”. The innovations of the developed RDNN include: 1) support the supervised and reinforcement learning algorithms; 2) support various DNN models including compressed models; 3) use hybrid digital and analog computing units to minimize energy consumption; and 4) using posture and behavior recognition system to verify the performance of RDNN engine.This subproject entitled “Reconfigurable Deep Neural Network Techniques for Supervised Learning” attempts to develop reconfigurable design techniques for DNN with supervised learning. The innovations of this subproject include: 1) reconfigurable technique for CNNs under the constraint of end-to-end latency; 2) reconfigurable technique for various CNN models; 3) reconfigurable technique for hybrid digital and analog computing units.
