Discover the research topics of our PhD students
Our first PhD students at KInIT began studying and researching their topics in September 2021. Some of the research topics are inspired by industry challenges, some by ideas from top scientists abroad or by international projects.
On March 9th 2022, we had the pleasure to welcome our mentors, partners, students and other virtual guests to our special PhD showcase, where we presented the work of seven PhD students at KInIT.
During this first part of the event, the doctoral students shortly presented their research and replication studies. Keep on reading for a brief introduction to each of the topics below.
Afterwards, the guests were invited to join seven separate online rooms with each of our PhD students to dive deeper into the topics.
Natural Language Processing: Probing Understanding of Multimodal Models
“Combining the language and image content is very natural for human beings, but it is a very challenging task to train deep learning models to do it. Vision-language multimodal processing has many applications, such as helping visually impaired people, understanding images posted on social media and banning the one against its policy. It can be also used to describe MRI or CT scans in health care.
However, neural network-based models have one disadvantage: they have millions of parameters, and they appear as a black box to us. In my work, I focus on the interesting research question, if the state-of-the-art multimodal transformers achieved the results on tasks thanks to a deep understanding of both modalities or based on wrong impressions.”
Misinformation detection: Adversarial Examples Generation with Human-in-the-loop
“False news detectors have the job of predicting if a given piece of information is either true or false. Adversarial samples, that are created using deceiving perturbations on the input text, are used to increase a model’s robustness. We collected these samples using a game, where players were producing such perturbations in order to trick a false news detector. This provided valuable insight on how to create such samples which is the basis for our future research.”
“I am working on off-policy evaluation and optimization of recommender systems. Basically, I am working on a better way to evaluate new algorithms than to use your customers as lab rats for A/B testing.
It improves the overall quality of recommender systems and prevents deploying bad algorithms with potentially harmful effects, such as misinformation spreading.
I chose this subject as my external mentor is one of the biggest experts in this domain. It combines all the things I love – machine learning, reinforcement learning and statistics. It is a hot topic with potentially huge impact, which could lead to a fulfilling career afterwards.
The potential use of my work impacts any development and training process of a new recommender system. Imagine having better movie recommendations on Netflix or better posts selected for you on Instagram.”
“My research focuses on the detection of anomalies, such as cyber-incidents, on computer networks using machine learning. The ability to detect these events on a network level is crucial in order to identify attacks in their early stages and thus protect the critical infrastructure of commercial or government stakeholders. As I’ve been in the computer networking and cybersecurity fields for more than a decade now, I decided to utilize my expertise in research and hopefully make the Internet a safer place one day.”
Security: Catch more Malware with Deep Clustering
“In my work, I study how deep clustering methods can help improve clustering of benign and malware executable programs.
This can help speed up the detection of new, never-before-seen malicious programs that can potentially cause damages in millions of dollars if they go undetected. The issue of early detection is a crucial one – as in the last year there was a new record high in the number of new unique malware samples.
I chose this line of research because it gives me the opportunity to work on real-life problems in cooperation with our industry cyber security partner, ESET, while being able to utilize my expertise in machine learning and artificial intelligence.”
“The topic of my research is winter road maintenance using new approaches to predictive modeling and optimization algorithms, with a focus on machine learning and artificial intelligence methods. Together with our industry partner, SOFTEC, we aim to make the winter commutes safer and faster for the drivers.
At the same time, we want to minimize the human labor and material costs and reduce the environmental impact of the maintenance services. Protecting the environment has always been very important to me, so the possibility of reducing the environmental damage was my biggest motivator to choose this topic.”
Misinformation detection: Stability of Machine Learning with Limited Labelled Data
“In my PhD studies, I am working on making the job of fact-checkers easier. I help to automate some parts of the fact-checking process by using machine learning as a part of the international research project CEDMO.
The domain of misinformation is characterized by the lack of labels spread across languages and tasks, so we employ approaches for machine learning with limited labelled data that can achieve good transferability. The success of the transfer and the overall stability of performance is affected by many factors. Therefore we investigate the factors that influence the stability and the success of transfer, in order to better control their effects, which will in turn help in development of more successful and efficient machine learning tools for automating tasks required for fact-checking.
This way, I am working on a current, real-life problem and thanks to the research project, I can quickly see the results of my work in practice.”
We would like to thank HubHub for the space that we could record in.
KInIT PhD program is run in partnership with the Faculty of information technology Brno University of Technology.
Research presented in this showcase was partially financed from the state budget by