Federated Self-supervised Learning in Computer Vision

Thesis Project owned by Scaleout Systems
1y ago update
With an ever-increasing amount of available image data, self-supervised learning (SSL) circumvents the necessity for annotations in traditional supervised learning methods. SSL methods such as SimSiam have shown excellent results on popular benchmark datasets, even outperforming supervised methods. Federated learning, on the other hand, is transforming traditional centralized training of machine learning models to distributed training on edge devices (e.g., phones, self-driving cars). Data stays on devices, and only model updates are shared with a central server. This reduces data migration costs and preserves data privacy. 

Even though edge devices collect massive amounts of image data, it cannot be labeled due to missing interfaces, expertise, or the reluctance of users to label their own private data. SSL is able to address this label deficiency. While SSL has achieved impressive results in settings where data is stored centrally, its performance in federated settings remains relatively unexplored. The present study investigates federated self-supervised learning (FSSL) in the domain of computer vision. First, four FSSL methods based on SimSiam are implemented and evaluated on CIFAR-10.

This work proposes FedAlign, which outperforms FedSimSiam in linear evaluation and achieves highest accuracies in IID settings. Results further show that Hetero-SSL [26] performs best in non-IID settings. Additionally, FSSL is applied to an autonomous driving dataset with the aim of detecting cars. The findings show that the weights obtained through FSSL are able to closely match supervised ImageNet weights, without the need for expensive annotations. The results highlight the potential of FSSL methods in real-
life applications.


Information Technology
IT & Software
More Efficient