Home › Categories › Business & Finance› Download

Download.pdf

improving-self-supervised-representation-learning-via-sequential-adversarial-masking.pdf

🔗 Source: assets.amazon.science

📊 Size: 734 KB

📄 Pages: 9 pages

⬇️ Downloads: 63

Summary

Recent self-supervised learning methods have shown that masking-based pretext tasks can be effective in computer vision, but existing approaches use random or ad hoc masking strategies that limit the difficulty of the reconstruction task. A new framework is proposed that generates masks in a sequential fashion with different constraints on the adversary, leading to improvements in performance on various downstream tasks, such as classification on ImageNet100, STL10, and CIFAR10/100 and segmentation on Pascal VOC. The sequential masking framework generates masks one by one, with a budget parameter controlling the strength of each mask and an overlap penalty encouraging the masks to capture different regions of the image. This approach achieves better performance than prior work, especially on linear readout on ImageNet100s, and comparable performance in fine-tuning and semantic segmentation on Pascal VOC.

Download File

Description

Recent self-supervised learning methods have shown that masking-based pretext tasks can be effective in computer vision, but existing approaches use random or...

Technical Information

File Format: PDF
File Size: 734 KB
Pages: 9
Language: EN
Total Downloads: 63
Last Updated: 1 week ago

Document Overview

This PDF document about Download provides comprehensive information and guidance. Whether you're a beginner or advanced user, this resource offers valuable insights into Download.