Seoul National University's College of Engineering (Dean Byoungho Lee) announced on October 29 (Friday) that a research team led by Professor Byung-Gon Chun of the Department of Computer Science and Engineering, jointly with FreindliAI, has developed a system called 'Terra' that allows all deep learning programs to be carried out more rapidly without modification.
 
With the development of Terra, it has been made possible to carry out symbolic graph processing and imperative processing at the same time, resulting in a much faster processing speed. This study proposes a performance model for a new deep learning program that did not exist before, suggesting a new direction for future deep learning framework technology.
 
Both PyTorch and TensorFlow, which are currently widely used deep learning frameworks, require the writing of deep learning programs, just like for ordinary Python programs. Like a general Python program, the written deep learning program is processed by the Python interpreter to perform artificial intelligence operations. However, the execution method using the Python interpreter has the disadvantage of a slow learning speed because there is no symbolic graph that expresses the entire deep learning operation at once, and therefore, it is impossible to attempt for overall optimization. To solve this problem, in previous studies, various methods have been proposed to convert a given Python deep learning program into a symbolic graph through Just-In-Time (JIT) compilation technique and to execute the program with a separate symbolic graph execution engine. However, the method using JIT compilation also had limitations in that the user had to modify the code, the accuracy of the program's execution would not be guaranteed, or there would be limited support for Python language.
 
Accordingly, Professor Jeon Byung-gon's research team and FriendliAI proposed a completely new Python deep learning program execution method. Instead of converting the entire Python program into a symbolic graph as in the conventional method, Terra collects only actual deep learning operations from the program and generates a corresponding symbolic graph for it. Then, it supports the execution of Python programs, which was not possible with the JIT compilation method, by simultaneously performing non-deep learning operations in the generated symbolic graph and a given Python program. Terra's simultaneous execution engine is the world's first method that guarantees the accuracy of program execution without requiring modification of the Python deep learning program written by the user and allows all syntax of the Python language to be used as it is. Terra provides deep learning acquisition speed of up to 1.73 times faster than TensorFlow's basic imperative execution method, and can even perform programs that the latest JIT compilation-based Autograph used in existing Tensorflow cannot perform, at an even faster speed. The results of this study will be presented at NeurIPS (Neural Information Processing Systems) 2021. “Terra: Imperative-Symbolic Co-Execution of Imperative Deep-Learning Programs”, Taebum Kim, Eunji Jeong, Geon-Woo Kim, Yunmo Koo, Sehoon Kim, Gyeong-In Yu, Byung-Gon Chun.
 
"I am delighted and honored to consistently announce world-leading AI platform technology. Recently, we are focusing on research and development of ultra-large AI, and FriendliAI is creating such ultra-large AI model development platform so that anyone can easily develop and use it," said Professor Byung-Gon Chun.