Online Semi-Supervised Learning and Face Recognition

This project focuses on real-time learning without explicit feedback. This work combines the ideas of semi-supervised learning on approximate graphs and online learning. In particular, we develop algorithms that iteratively build a graphical representation of the world and update it on-the-fly with observed examples (both labeled and unlabeled). We proved regret bounds of the solutions, demonstrated that the system can recognize faces in real-time even in a resource constraint environment and can take advantage of the manifold structure to outperform existing methods. The following videos show how online semi-supervised learning can be used to train a robust face recognizer of a person from just a single frontal image:


  • Online Semi-Supervised Learning on Quantized Graphs

    Michal Valko, Branislav Kveton, Daniel Ting, Ling Huang. In Proceedings of the 26th Conference on Uncertainty in Artificial Intelligence (UAI) , July 2010. [pdf]
  • Online Semi-Supervised Perception: Real-Time Learning without Explicit Feedback

    Branislav Kveton, Michal Valko, Matthai Philipose, Ling Huang. In Proceedings of the 4th IEEE Online Learning for Computer Vision Workshop (OLCV) , 2010. [pdf] Awarded best paper!
  • Semi-Supervised Learning with Max-Margin Graph Cuts

    Branislav Kveton, Michal Valko, Ali Rahimi, Ling Huang. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics (AISTATS), May 2010. [pdf]

Project Team