Like the year before, I was involved in the MSc. course Deep Learning, focusing on the supervision of student groups with their reproduction projects.
Reproduction projects
Empirical research should be reproducible such that it can be fact-checked and readily used for successive research. For deep learning specifically this means 1) a properly documented and working online-available code base and 2) an accurate indication of both the model architecture and optimization settings (preferably mentioned in the published article). Unfortunately, this is not always the case. In this course, the reproducibility of already published articles is exploited: small groups of students chose articles they prefer and reproduce (part of) its empirical results. Among the list of this year, we added several well known articles that are relevant for the lab’s research endeavours and supervised several groups of students for various topics: autonomous design of floorplans [1], self-supervised analysis of structural layouts [2], attention-based deep learning [3], and state-of-the-art contrastive learning [4]. Below, some of the blogs can be found:
- HouseGAN reproduction (1) [1]
- HouseGAN reproduction (2) [1]
- LayoutGMN reproduction [2]
- Attention-based deep metric learning reproduction [3]
- SimCLR(1) reproduction [4]
- SimCLR(2) reproduction [4]
References
- [1] House-GAN: Relational Generative Adversarial Networks for Graph-constrained House Layout Generation
- [2] LayoutGMN: Neural Graph Matching for Structural Layout Similarity
- [3] Attention-based Ensemble for Deep Metric Learning
- [4] A Simple Framewok for Contrastive Learning of Visual Representations
Overview of all reproduction projects? Click here.