Defence of doctoral thesis in the field of computer science, Vikas Verma

How to train large scale deep neural networks with fewer labeled examples?
CS_defence_3 photo by Matti Ahlgren

Title of the doctoral thesis is "Algorithms for Data-Efficient Training of Deep Neural Networks"

State-of-the-art deep neural networks have had a transformative impact across many tasks, yet this improvement typically can only be achieved by using hundreds of thousands (ideally millions) of labeled data samples. This is due to the fact that modern-day large scale deep neural networks contain millions of trainable parameters. In many tasks, collecting this labeled data is either difficult or impossible. For example, segmented medical images can be produced by a skilled human annotator, yet doing this for a significant number of images is very expensive and time consuming.

In this thesis I have explored a class of methods which improve the data efficiency of DNNs by constructing synthetic samples. These synthetic samples are constructed by combining two or more real samples. Depending on the problem setting, e.g., fully-supervised learning, semi-supervised learning, adversarial learning, and unsupervised learning, this thesis proposes various novel methods for leveraging such synthetic samples. In addition, this thesis also explores methods for training graph neural networks, which are an important class of neural networks for domains where samples naturally exist as graphs, e.g. in drug discovery. The methods proposed in this thesis are easily reproducible and achieve state-of-the-performances across many problem settings. Furthermore, these methods have already been deployed by other research labs in many data-critical applications, such as for medical image segmentation.

Opponent: Professor José Miguel Hernández Lobato, University of Cambridge, United Kingdom

Custos: Professor Juho Kannala, Aalto University School of Science, Department of Computer Science

Contact information of the doctoral candidate: Vikas Verma, Aalto University School of Science, Department of Computer Science, [email protected] and +358 50 4685 953

The defence will be organized via remote technology (Zoom). Link to the defence.

Zoom Quick Guide (

The doctoral thesis is publicly displayed 10 days before the defence in the publication archive of Aalto University (

Electronic doctoral thesis (

  • Published:
  • Updated:
URL copied!