In this thesis, you should explore different approaches to learn writer indepen-dent digit representations and compare them with each other.
Digit recognition on the MNIST dataset is the ”Hello World” of deep learning.This is the case, since it is easily solved using simple convolutional neural net-works (CNNs). However, the implicit assumption of the produced model is that the digits it will be used on are written in a similar style. If this assumption is violated, one can expect the digit recognition performance to drop. One possible way to avoid the impact of this limitation is to learn representations of the given input images, which are writer independent. This approach is followed, for example, by Zhang et al. , who use adversarial training to learn digit representations, which are the same for printed and handwritten digits.
In this thesis, you should explore different approaches to learn writer independent digit representations and compare them with each other. Possible initial directions would be the adversarial training approach by Zhang et al. , as well as other approaches aimed at learning robust representations, such as triplet networks  or capsule networks .
Experience in Python Programming
If you are interested in this topic, please contact me via e-mail or come to my office.
Florian Westphal e-mail: firstname.lastname@example.org office: J3119
 Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International Workshop on Similarity-Based Pattern Recognition, pages 84–92. Springer, 2015.
 Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. In Advances in Neural Information Processing Systems,pages 3856–3866, 2017.
 Yaping Zhang, Shan Liang, Shuai Nie, Wenju Liu, and Shouye Peng. Ro-bust offline handwritten character recognition through exploring writer-independent features under the guidance of printed data. Pattern Recognition Letters, 106:20–26, 2018.