計畫團隊成員 Members

 
 
圖片1-2

總計畫暨子計畫一
Main Project & Subproject 1

林永隆  清華講座教授

國立清華大學資工系

圖片2-2

子計畫二
Subproject 2

吳凱強  助理教授

國立交通大學資工系

圖片3-2

子計畫三
Subproject 3

黃俊達  教授

國立交通大學電子所

 
 
 
圖片4-2

子計畫四
Subproject 4

黃世旭  教授

中原大學電子系

圖片5-2

子計畫五
Subproject 5

鄭維凱  副教授

中原大學資工系

圖片6-2

子計畫六
Subproject 6

王廷基  教授

國立清華大學資工系

 
 

技術亮點 Technical Highlights


人工智慧邊緣運算的硬體加速解決方案:
- 支援多種神經網路(DNN, RNN, GRU, LSTM)的硬體加速器
- 支援多種神經網路(DNN, RNN, GRU, LSTM)的模型壓縮技術
- 軟體開發套件
- 語音指令辨識實作展示
高解析度即時影像語義分割技術:
針對高解析度(1024*2048)視訊,語義分割達到80 fps的推論速度
即時物件辨識技術 
創新卷積運算架構與近似計算乘加單元
對抗式攻擊與防守 記憶體內運算

Hardware accelerating solution for AI on edge
- Hardware accelerator supporting DNN, RNN, GRU, LSTM
- Neural network compression supporting DNN, RNN, GRU, LSTM
- Software development kit
Novel image semantic segmentation achieving 2K*1K resolution@80 fps
Real-time object detection
Novel convolution architecture and approximate MAC computing
Adversarial attack and defense
In-memory Computing

 

應用情境 Applications


低功耗高性能AI硬體加速器可廣泛應用於IC設計、電子資通訊、交通運輸、家電、消費性電子、健康照護等產業。 高解析度即時影像分割與辨識技術可應用於自動駕駛、醫療診斷、安全識別、人機互動等產業。

Low-power deep learning accelerator can be used in IC design, communications, transportation, home appliances, consumer electronics, e-health etc. related industry. The image semantic segmentation and object detection technology can be used in autonomous driving, medical diagnosis, security surveillance, human-computer-interface, etc.

 

其他技術介紹


語音指令辨識系統之即時對抗式攻擊技術
常見的對抗式攻擊方法主要是以疊代方式產生對抗例,雖然有極高的成功率,但執行上也相對耗時。針對語音指令辨識系統,本計畫提出創新的對抗式攻擊技術,透過事先訓練好的遞迴類神經網路,能夠即時產生對抗例。與目前已知的先進方法相較,我們的技術在執行速度方面快了400倍以上,同時具有相近的攻擊成功率。

Real-Time Adversarial Attack for a Keyword Spotting System
Previous methods of performing adversarial attacks against speech recognition systems often treat this problem as an optimization problem and require iterative updates to generate optimal solutions. Although they can achieve high success rate, the process is too computational heavy even with the help of GPU. In this project, we develop a novel real-time adversarial attack method, which applies recurrent neural network with two steps of training processes to generate adversarial examples targeting a keyword spotting system. The empirical studies show that the execution time of our attack is more than 400 times faster than a state-of-the-art attack (i.e., C&W attack) with the comparable attack success rate.

 

影音介紹 Media