How to win an AI Hackathon?
Get to know the winning team of the Hackathon spring 2021 hosted by BRAIN x COGITO.
The team:
Anders: I am a 4th-year computer science student at NTNU, focusing on Algorithms and High-Performance Computing. Besides having completed a bachelor’s degree in Computer Science, I have also been a part of the startup culture most of my time studying. Most of my focus has been software development, which has given me real-life experience in the field. At the time of writing, I am a co-founder and project lead in Favn Software, where I am working with Dilawar and William from this hackathon. Most of my passion lies in software architecture and databases, but I am interested in machine learning and AI. In my spare time, I focus on socializing, working out, and playing guitar.
Dilawar: My name is Dilawar Mahmood, and I’m 21 years old. Currently, I’m pursuing a bachelor’s degree in Computer Science at NTNU, where I’m writing about privacy-preserving federated learning on decentralized medical data. In addition to my studies, I’m also taking extra courses in AI, statistical learning, abstract algebra, and software engineering. After finishing my bachelor’s degree, I plan to take a master’s degree in artificial intelligence. I also work at two startups in Trondheim, where I do software development and data science. In my spare time, I enjoy competitive programming, doing software engineering and AI projects, working out, watching movies, and hanging out with my friends.
Helle: Hi, my name is Helle. I study Computer Science at NTNU, where I’m currently writing my master thesis on semi-supervised artificial intelligence. In addition to my studies within the field of AI, I’m pursuing a bachelor’s degree in social economics and a bachelor’s in theoretical physics. I also work part-time at an IT-consultant company here in Trondheim, where I get to work on a range of exciting projects. I do, of course, have some interests that are not related to school, such as cooking, makeup, music, and wine-and-sushi nights with friends.
Simen: I’m Simen, and I discovered programming a decade ago (yes, writing that made me feel really old). I’ve worked on a gazillion projects, from extending the lifetime of a million phones by backporting OS features, to launching new functionality on one of Norway’s most visited websites. Currently, I’m writing my CS master with Telenor and NTNU, incorporating the latest research on neural language models into Telenor’s research on speech recognition for low-resource languages. Whenever I’m not deep-diving into some project, you’ll usually find me hiking somewhere, playing board games, or just enjoying a good movie.
William: My name is William Chakroun Jacobsen. I am from Oslo, Norway, and I’m currently studying Computer Science with a specialization in Algorithms and HPC. I have completed a bachelors’ degree in Software Engineering, where I wrote my thesis about Interpreting EEG signals using Deep learning. In this thesis, I used an ensemble of models. In my spare time, I work with a Startup in Trondheim named Favn Software with Anders and Dilawar. At Favn, I work as a project Lead where I have been working with various projects, such as App development and architectures to store data. I do also relax in my spare time. I play D&D with my friends, watch movies, and workout once in a blue moon.
Where we found the event? Facebook / Brain people
How did we approach and solve the problem?
We chose the unsupervised anomaly detection task. Our main approach was divide and conquer with peer programming on different tasks. Since we were five members, it was easy and efficient to work in that way. We split the main task into several subtasks, which involved:
- Data preparation and feature engineering
- Model development
- Model evaluation
- Visualization
- Final presentation
Everybody on the team had experience with the data science and machine learning tools we used in the hackathon, so all the team members got to work on the different subtasks listed above.
We also had regular “check-ins.” Even though we were in the same room, we regularly took short meetings to find out what our status was and what our following priorities were. The latter included finding out which AI models to continue working on and any other work that needed to be done to fulfill the requirements provided by Telenor.
We started with building a simple baseline model, which used standard deviations to find anomalies. After that, we developed two somewhat complex models. The first one was an isolation forest that isolates anomalies instead of the most common techniques of profiling normal points. The other model was a vector autoregressive model (VAR); a statistical model used to capture the relationship between multiple quantities as they change over time. We ended up removing the latter model, as isolation forests were better at detecting anomalies.
We also trained a DeepAnT model. It is an unsupervised time-based anomaly detection model, which uses convolutional neural network layers to keep the number of trainable weights at a minimum (so it doesn’t overfit) while still being able to build good internal representations of the data. We found that it was able to make use of correlations between the features to make far better detections than our other models.
Together, this ensemble of AI models gave us a recall rate of around 98% on the test data provided by Telenor. As this was a truly unsupervised problem, precision could not be measured accurately, so instead, we carefully studied far too many visualizations to verify that we had not sacrificed too much precision to achieve this recall rate.
What do we think about the event?
We enjoyed working with a real-world dataset, where the task was open enough for us to explore different methods concerning feature engineering and AI models. Even though we were only able to test out a subset of possible solutions, we believe that there is a lot of potential for using different AI-techniques on these types of problems.
We think that the digital hackathon was pretty relaxing. We could eat and work when we wanted to and controlled our own time, even though we ended up working till 4 AM. Food from Foodora was delicious.
