Professor Hyun Oh Song of the SNU Department of Computer Science and Engineering Presents a New Data Augmentation Technique to Significantly Improve the Practical Applicability of Artificial Neural Networks



▲ From left Professor Hyun Oh Song, MS/PhD Students Wonho Choo, Jang-Hyun Kim, Hosan Jeong

 
■ Development of algorithms to optimize the random elements of traditional data augmentation techniques
■ Creation of new data that is efficient for learning with limited learning data
■ Improved performance of artificial neural networks such as object recognition, location detection, and robustness
 
Professor Hyun Oh Song and his research team (MS/PhD Students Jang-Hyun Kim, Wonho Choo, and Hosan Jeong) have developed Co-Mixup, a data augmentation technique that improves the general performance and robustness of artificial neural networks.
 
Specifically, when given a large amount of data, Co-Mixup aims to maximize the amount of important information in each data while at the same time, generating various data [Figure 1]. To achieve this, the researchers designed a formal problem and developed an algorithm to solve it effectively. [Equation 1].
 
Co-Mixup will be presented at an oral session of the International Conference on Learning Representations (ICLR), one of the top AI conferences, that is to be held online in May this year. The session is a publication opportunity given only to the top 1% (=53/2997) of papers that were submitted to ICLR this year.
 

Picture 1. Example of Co-Mixup and comparison with conventional methods

 

Equation 1. Co-Mixup's objective function that measures the amount and variety of information in the generated data
 
References
Jang-Hyun Kim, Wonho Choo, Hosan Jeong, Hyun Oh Song, Co-Mixup: Saliency Guided Joint Mixup with Supermodular Diversity, ICLR 2021
https://openreview.net/forum?id=gvxJzw8kW4b