General Questions
What is the Norwegian AI Championship?
The Norwegian AI Championship is a one-week hackathon-style competition running from August 1-8. Organised by NORA.ai (Norwegian AI Research Consortium) in Norway, it is the first-ever nation-wide championship that will engage bachelor and master students in AI. The championship is in partnership with the Danish AI Championship (now in its 5th year) and uses the same infrastructure, challenges, and jury system. Norwegian teams compete among themselves for the title of "National Champions of AI in Norway," while Danish teams compete separately for their own title.
Who can participate?
Bachelor and master students across Norway. Participation can be as part of a team or individual. Register to participate on this link.
What's the maximum number of members each team can have?
No cap set, but 2-5 people is recommended. Solo participation is possible (had solo runner-up last year), but teams of 2-5 have been most successful.
What community resources can I access?
Discord channel: https://discord.gg/QVSZECgw8g
(great place to connect with others and/or find team members)
Information webinar: Youtube Link (coming soon)
Advice from last year’s winners: https://norwegian-ai-championships-guide.lovable.app
Physical Communities:
A* consulting (Trondheim)
Cogito NTNU: will be hosting in-person sessions where students will get together to code, if you want to join them contact:
Afras Mansoor: afrasum@stud.ntnu.no
BETA student group (Agder)
What is the objective of the championship?
Competition format: Solve 3 different AI use cases, each worth 1/3 of total points
Technical goal: Build full end-to-end machine learning systems with API endpoints (not just models)
Evaluation method: Performance-based scoring using well-defined metrics for each use case
Broader purpose: Create a community of like-minded AI students, encourage real-world application building, and push boundaries of what students can achieve
What criteria do judges focus on most when evaluating submissions?
A: Model performance based on well-defined scoring metrics. Everything is automated based on algorithm performance. However, code robustness is also considered if issues arise during competition.
Technical Questions
Can I use external data?
Yes, absolutely encouraged! This is highlighted as a key differentiator:
Winning strategy: Last year's winners found additional medical images online for their CT scan task
Data modification: You can synthetically modify external data to match competition data distribution
Examples: For sentiment analysis tasks, you can gather ratings from different websites
Competitive advantage: Teams using only provided data vs. teams gathering additional data - external data often gives significant advantage
Important: Ensure external data matches the target distribution you're trying to model
What programming language can I use?
There is no specification on a programming language.
Python: Clearly the primary language (references to try-except statements, Python libraries)
API development: Must be able to create API endpoints that can be queried
Performance languages: For some tasks, faster languages like C++ or Rust mentioned as potentially beneficial, but only if absolutely necessary
Framework flexibility: Can use any frameworks/libraries (Hugging Face models frequently mentioned)
What computational resources are available?
You can apply for student computing quota either with your local university, Google, Amazon or Azure 💻
Are there cloud services like Azure to support training large models?
Norway working on solutions - check with NORA for specific Norwegian resources.
Stucturing Your Time
How much time has each winning team used per championship?
Significant hours invested. Need to optimize between sleep deprivation and productivity. It is not just about hours, but quality hours when mentally available.
Are we allowed to start working before August 1st?
You can't work on actual challenges (not revealed until kickoff), but you can do preparatory work like learning API setup, server communication, etc.
Questions for Last Year’s Winners
How long did pre-training take on 8x RTX 4090 GPUs?
Around 24 hours of training on internet data, with multiple training rounds due to errors. Strategy: run longer training overnight while sleeping.
What about batch size?
No magical batch size - tried different ones. Recommended reading what others have done with similar models on Hugging Face.
How difficult was finding and gathering data?
Took several hours to find, download, extract, pre-process and upload to GPU server. Process involves scouring internet for relevant data sources.
Did you try other models than U-Net? What's the strategy to pick good models?
Yes, tried other models and U-Net variations. Strategy: Split training data into train/validation sets, train different models on training data, validate on validation data, pick best performer. Crucial to pick models meant for your specific task.
How long did API deployment take?
About an hour, especially easy with ngrok.
Is the data structured or do we need to pre-process it?
A: Depends on task. Generally not completely clean/easy-to-use. You get basic X/Y structure for supervised learning, but may need preprocessing. Additional data gathering requires own structuring and labeling.