Master Thesis - Semi-supervised learning
Master Thesis & Research Opportunities - Gothenburg, Sweden
The most common way of training a deep neural network is supervised learning, where the network is trained to mimic the ground truth, usually labels supplied by a human annotator. However, only a tiny fraction of the data gathered by Zenuity's data collection fleet can be labelled in this way. There is therefore a lot of potential gain from making use of also the unlabelled data in training. This is called semi-supervised learning.
At Zenuity, we use deep neural networks to solve a number of perception tasks, such as semantic segmentation and object detection, in 2D and 3D. The goal of this thesis project is to improve the accuracy of one or more of these tasks by making use of unlabelled data in the training process. At a conceptual level, this can be achieved by training the networks to be accurate for labelled data, while also being consistent on unlabelled data. For example, the predictions of a good object detection network should be relatively stable across time.
In this master thesis project, you will focus on:
- Surveying the literature on semi-supervised learning from video
- Implementing and extending the most promising methods
- Making use of these methods to train deep neural networks on both Zenuity's labelled and unlabelled data
- Investigating how much the accuracy of these networks improve as a result
We are looking for two students, preferably with good knowledge of
- Python programming
- Deep learning
- Handling large amounts of data
Final application date: 2019-11-15. But please submit your application as soon as possible, as we will be screening candidates continuously.
Please send in individual applications with CV, motivational letter and grade transcripts. If you wish to partner with someone, simply note that in your application.
Planned start: January-March 2020.
Duration: 30 ECTS
For questions regarding the project, please Contact: email@example.com.