| 摘要: | 目前有關深度學習神經網路引擎之電路實現的研究可以分兩大類：數位電路實現及超低功耗類比電路實現 (譬如: 仿神經型態運算)。為了尋求達到更多耗能之下降，也有一些研究提出以新興技術來實現DNN的作法，參考文獻顯示，於物體辨識應用時其能源效率比RISC軟體實現的能源效率高約十幾萬倍以上，比CMOS數位電路實現方式的能源效率高百倍以上，且所需面積僅是CMOS數位電路實現方式約0.16倍。隨著電子設備尺寸的不斷縮小，最新的計算技術進步使得可穿戴設備的設計可以實現長期連續監測的任務，並有可能促進及時的醫療措施進行治療和護理，這就是可穿戴設備吸引了該領域科學家廣泛關注的原因。在生醫訊號中，醫生經常會採用心電圖(Electrocardiogram, ECG或EKG)與心音圖(Phonocardiogram, PCG)作為判斷心臟疾病的參考。此外，隨著近年來科技的演進，系統能處理更大量的運算使人工智慧(Artificial Intelligence, AI)再度成為研究熱門。將人工智慧應用於語音處理，透過以Deep Neural Network (DNN)為基底的應用，使自動語音辨識(Automatic Speech Recognition, ASR)更廣更精準的進步。本計畫主要研究方向是針對仿神經型態運算技術開發高能源效率可重組架構及晶片電路，將分年著重基本神經元相關電路架構設計及規格釐訂、仿神經型態運算架構及模組開發、可擴充及可重組架構開發，並且於每一階段皆將進行晶片設計與試製來加以驗證。本計畫將提出一個神經細胞(Neuron Cell)模型，組成樹突神經模型(Dendritic Neuron Model, DNM)架構並以粒子群最佳化演算法(Particle Swarm Optimization, PSO)，實現深度學習用於ECG/PCG辨識及語音辨識系統，以上述應用作為測試檢驗平台來展示研發成果。 ;The current research on the circuit implementation of deep learning neural network engines can be divided into two categories: digital circuit implementation and ultra-low-power analog circuit implementation (ex., neuromorphic computing). In order to achieve more reduction in energy consumption, some studies have also proposed the use of emerging technologies to implement DNNs. The literature shows that the energy efficiency of analog neuromorphic computing for object recognition applications is about 100,000 times higher than that achieved by RISC software. The energy efficiency of analog neuromorphic computing is more than a hundred times higher than that of the CMOS digital circuit implementation, and the required area is only about 0.16 times that of the CMOS digital circuit implementation.The computer can process more complicated system due to the evolution of science and technology in recent years, and it also making the Artificial Intelligence (AI) get a second wind. As electronic devices continue to shrink in size, the latest advances in computing technology have enabled the design of wearable devices to achieve long-term continuous monitoring tasks, and have the potential to promote timely medical measures for treatment and care. In the biomedical signal, doctors often use Electrocardiogram (ECG) and Phonocardiogram (PCG) as the reference for judging heart disease. In addition, with the evolution of science and technology in recent years, the system can handle a larger number of operations, making Artificial Intelligence (AI) once again a research hotspot. Applying artificial intelligence to speech processing includes significantly reducing the Word Error Rate (WER) of speech recognition. It will improve automatic speech recognition (ASR) for more accurate and wide range by the applications through the Deep Neural Network (DNN).The main research direction of this project is to develop high-energy-efficient reconfigurable architecture and chip circuits for neuromorphic computing technology. It will focus on the design and specifications of basic neuron-related circuit architecture, neuromorphic computing architecture and modules. This project will propose a neuron cell model, which will form a Dendritic Neuron Model (DNM) architecture and use Particle Swarm Optimization (PSO) to implement deep learning for ECG/PCG recognition and speech recognition system, using the above-mentioned applications as a test and inspection platform to show the results of research and development. |
