T·01
Fundamentals
SVMs, logistic regression, decision trees, kernels: the classical core, derived by hand. This is the math everything else stands on.
The IMLC tests how deeply you understand AI: derive the math behind SVMs and gradient descent, reason through experiments without touching a keyboard, and read the research shaping LLMs, vision models and AI security. Closed-book. Three rounds, one leaderboard.
This is not another Kaggle. There are no GPUs to burn and no ensembles to stack. Every round is closed-book, and what gets tested is understanding: theory you can derive, experiments you can reason through, research you can actually read.
Margin maximization, gradient descent, attention: you derive it and defend it instead of importing it from a library.
No running code, pseudocode at most. But dev splits, leakage, overfitting and ablations are all fair game: the judgment behind good experiments, tested as understanding.
In the pre-final round you receive a current paper on topics like RAG, adversarial attacks or efficient inference, and answer questions that prove you understood it.
P3.Two gradient-descent runs, A and B, descend the training loss (left). The right panel shows the same two runs on the test-error surface.
(a)Contour bands are tighter above the basin than to its left. At which start, A or B, is ‖∇L‖ larger?
(b)Both runs end at the train-loss minimum, yet not at the test-error minimum. What does this gap illustrate, and why does it favour wide, flat minima over sharp ones?
(c)The step dots bunch up as the runs approach the minimum. Why, when the learning rate is fixed?
Six areas, one spectrum: the fundamentals that never change and the research that changes every month. Every round draws on all of it.
SVMs, logistic regression, decision trees, kernels: the classical core, derived by hand. This is the math everything else stands on.
How machines actually learn: gradient descent, loss landscapes, regularization, and why models that merely memorize fail to generalize.
Backpropagation by hand, then the architectures that took over: CNNs, RNNs and the transformer, understood layer by layer.
The current research edge: LLMs, fine-tuning, RAG, and the efficiency work, from quantization to distillation, that makes them deployable.
Where it all lands: vision, healthcare, recommender systems, science. The same methods, translated into systems people rely on.
Adversarial attacks, robustness, alignment and fairness: how models fail, how they get attacked, and what responsible deployment demands.
Each round raises the stakes: first the fundamentals, then a real research paper, then one final exam where everything converges.
Closed-book fundamentals: derivations, conceptual questions and pseudocode covering SVMs, logistic regression, gradient descent and the judgment calls behind good experiments. Reach 15/17/20 points (junior/youth/senior) to advance.
You receive a recent research paper on a topic like LLM efficiency, vision transformers or adversarial robustness, and answer questions that test whether you truly understood its method, its assumptions and its limits. You have five days. Registration: 12 EUR.
One closed-book exam of 30 questions, against the clock. Fundamentals, modern architectures and research judgment converge, with no compiler between you and the answer. Results announced 23 January 2027.
Solutions for the qualification round are due by Friday, 6 November 2026.
no GPUs required · open to students worldwide · financial aid available
Register your students, follow their progress, and put your school on the leaderboard. Partner schools get reduced registration for every participant.
Manage and track your students through every round, from qualification to the final exam, in one dashboard.
Register as a teacher →Partner schools receive reduced registration fees for all students and a certificate shipment to the school.
Become a partner school →Posters, problem sheets and training material to introduce machine learning thinking to your students.
Get the materials ↓Ambassadors bring the competition to their schools, universities and communities. You spread the word, mentor newcomers, and grow a local ML community — we support you with materials, a certificate and an international network.
Share the competition at your school or university and on local channels — every registration through you counts toward your ambassador level.
Help participants in your region with questions about rounds, registration and preparation.
Advance through ambassador levels, earn certificates and references, and join an international network of ML enthusiasts.