발간년도 : [2019]
| 논문정보 |
|
| 논문명(한글) |
[Vol.14, No.5] Effective Implementation for Fast Deep Learning Algorithm |
|
| 논문투고자 |
Sangmin Suh |
|
| 논문내용 |
AI (Artificial Intelligence) based on deep learning has been successful in many application areas. Supervised learning such as image classification and object detection has been mainly used for vision and ADAS (Advanced Driver Assistance Systems) / AD (Autonomous Driving). And reinforce learning has been generally utilized for robotics and energy optimization. Therefore, in order to improve the performance, many research papers have focused on optimizing neural networks. However, in practice, FPS (frame per second) is a hidden and critical factor because FPS is also included in the performance measurement. This note show that pre-processing and post-processing are major components affecting FPS. And It is verified that FPS cannot be improved by optimizing the neural network itself because the pre-processing and post-processing are out of the neural networks. In this note, fast pre-processing methods on the basis of DSP (digital signal processing) is suggested. For DSP implementation, binary arithmetic is presented and quantization error due to the conversion from floating point calculation to fixed point calculation is discussed. In addition, major design frameworks for deep learning algorithm implementation are compared and their merit and demerit are also summarized. In the note, implementation is categorized into three, i.e., input data generation with pre-processing, model design of neural network, and performance evaluation. With the selected framework, detailed implementation is also presented. |
|
| 첨부논문 |
|
|
|
|
|
자료실 > 논문지 
