Seoul National University's College of Engineering (Dean Kookheon Char) announced on May 12 that Professor Byung-Gon Chun's team of the Department of Computer Science and Engineering has developed a Revamper system that optimizes the data augmentation process to perform machine learning at up to twice the speed of existing systems. It is expected that the system will enable more efficient AI learning in various fields.
 
Data augmentation refers to increasing the number of practical learning data by applying arbitrary transformation operations to the data that is being learned. Although data augmentation increases the accuracy of artificial intelligence learning models, it has the problem of slowing down the speed of its learning.
 
To address the problem of slow learning speed, Professor Byung-Gon Chun's team developed a novel data caching [1] system, Revamper. Existing methods proposed by Google speeds up learning by reusing the final augmented sample a certain number of times, but the accuracy of the model is compromised.
 
On the other hand, Professor Byung-Gon Chun's team proposed a data refurbishing technique that reuses samples without compromising the accuracy of the learned model. Data refurbishment solves the model accuracy degradation problem by dividing the data augmentation process into two parts and reusing samples with partial data augmentation operations for a certain number of times while performing the remaining augmentation operations before being used for training.
 
The researchers implemented a new caching system, Revamper, that uses reusable samples evenly across multiple learning steps to efficiently support the refurbishing scheme. Revampers provide AI learning speed of up to 2x faster than when compared to PyTorch [2] data loaders. The designing of Revamper was made in consideration of the convenience of its user and the existing pytorch model can be performed quickly using the Revamper. The researchers plan to unveil Revamper to allow pytorch users to utilize them.

The findings will be published on July at the USENIX Annual Technical Conference (ATC), a prestigious conference on computer systems.

"I am happy to consecutively announce a world leading artificial intelligence platform technology. From now on, we will create a super-large artificial intelligence system through friendli.ai and provide it as a service," said Professor Byung-Gon Chun while commenting on its development.
 
 
[Title of the research paper]
“Refurbish Your Training Data: Reusing Partially Augmented Samples for Faster Deep Neural Network Training”, Gyewon Lee, Irene Lee (Georgia Institute of Technology), Hyeonmin Ha, Kyunggeun Lee, Hwarim Hyun, Ahnjae Shin, and Byung-Gon Chun.
  [1] A method to improve processing power by storing the generated data and providing it quickly when requested
[2] Open source machine learning framework