[한국인공지능학회] Data Augmentations Affecting Downstream Performances Exploring the Impact of Gaussian Blur in Self-Supervised Learning - Sanskriti, Seungwoo Jang, Dongwon Kim, Kwangsu Kim
- 인공지능융합 연구실
- 조회수249
- 2023-07-20
Abstract
This paper explores the influence of Gaussian blur parameter settings on Self-Supervised Learning (SSL) models. By fine-tuning the MoCo[7] model with varying levels of Gaus- sian blur, we investigate its effect on downstream classification accuracy. Our experiments reveal that minimal to no Gaussian blur maintains acceptable performance, enabling the model to preserve subtle features for classification tasks. However, increasing blur levels result in a gradual decline in accuracy, highlighting the sensitivity of SSL models towards blur. The findings emphasize the critical role of parameter selection in achieving an optimal balance between capturing robust features and preserving essential details. By shedding light on the impact of Gaussian blur, this study potentially contributes to the significance of exploring the impact of various data augmentations used in SSL.