Multi-task deep learning is about learning multiple tasks using shared representations in a supervisory environment. Existing multi-task learning methodologies rely on splitting the network at a particular layer depending on the task. They do not generalize well across various tasks. We explore the split architecture and cross-stitch unit on facial landmark dataset where we regularize weights to share representation by introducing a mutual weight regularization loss.

Write-up

Poster

Citation:

• https://github.com/luoyetx/deep-landmark
• https://github.com/JeremyCCHsu/Kaggle_Facial_Keypoint
• https://arxiv.org/abs/1604.03539

Data:

• http://mmlab.ie.cuhk.edu.hk/projects/TCDCN.html
• http://mmlab.ie.cuhk.edu.hk/archive/CNN_FacePoint.htm

Version:

• Python v. 2.7.9
• Tensorflow v. 0.11.0rc1