Student Stories: Cohort 2 student, Kostas Alexandridis discusses object detection and his route to PhD
The Distributed Algorithms CDT (DA CDT) at the University of Liverpool engages its students in impactful, cutting-edge and diverse research, allowing them to collaborate and develop their skills within the university’s data science community. Read Kostas's journey with the DA CDT so far…
My Journey
I didn’t know that I wanted to do a PhD until I started writing my master's thesis in Electrical Engineering School at Democritus University of Thrace, Greece. I enjoyed going to the lab, doing my experiments and talking to other PhD students about research.
After completing my thesis, I started looking for a PhD in data science. At the same time, I was looking for applications that stood out from other PhD projects, like applications where machine learning can be used in scale. When I discovered the project ‘Drone Detect and Desist’ at the Distributed Algorithms CDT, I thought that it was an excellent opportunity. I liked the research proposal and the fact that there was industrial collaboration. I’m partnered with Vision4ce.
It was a challenge for me to apply for this position, as it was abroad, but it was also intriguing and exciting. In the end, I’m happy that I travelled to Liverpool and met new people that undertake brilliant research at the CDT.
My Project
At the moment I am working on Imbalanced Object Detection. This problem is encountered in real world applications, where the objective is to detect and classify all objects that a camera can see. The real problem in this domain, is that objects in the physical world follow a Zipfean Distribution. That means that there are objects that appear frequently and others that are more unique.
When one trains a deep neural network to detect such objects, then this autonomous system will do an excellent job of detecting the frequent objects but it will perform poorly in detecting the unique objects. This problem is called class imbalance and it is a serious problem that when it is not addressed properly, the autonomous system will be unreliable. For example, imagine a model that was trained with images inside a city. If this model is deployed to an autonomous car, then it will do an outstanding job detecting all the objects inside the city, but when this car is used in another environment with many animals, then this model may not be able to detect them and it may cause an accident.
Some techniques to solve the imbalance problem is data resampling, where essentially one reuses the images of unique objects during training. Another technique is cost sensitive learning, where one assigns more weight to training samples of unique objects.
In my case, during my studies I have developed two methods. The first is a cost sensitive learning method and the second is a method that replaces the activation function of the neural network with another activation function that enables rare category learning.
My Experience at the CDT
It is an amazing experience studying in the CDT. Every day I can go to the lab and talk with my colleagues, exchange new ideas, learn new things and undertake new experiments. Also, I have made new friends in the office and I enjoy it when we go out together to the university’s pub on Fridays.
Mini-Project
During my studies I did many side-projects. In my first year, I participated with fellow students in the High-Performance Computing (HPC) national student competition, where our team for University of Liverpool came second after completing a series of HPC challenges. Another side-project was the participation in the Siemens AI European student competition, where my project was admitted to the finals.
Distributed Algorithms CDT
The Distributed Algorithms CDT is an Innovative Data Science, AI and Machine Learning Research Centre, aligning PhD students, academics and industrialists to work together to generate novel solutions to tough data science challenges. If you would like to find out more about our programme and would like to talk about becoming an active member of our CDT community, please visit our website or email kelli.cassidy@liverpool.ac.uk.