​

The Syllabus, Broken Down

This is the section most families actually need — a plain-English map of what's actually tested, not just a list of buzzwords.

Math foundations

USAAIO doesn't test math in the abstract. It tests exactly the math that underlies the ML algorithms in the rest of the syllabus, and it expects students to derive results, not just recognize formulas.

​

• Linear algebra: vectors, matrices, and eigenvalues -  directly relevant to Principal Component Analysis (PCA), which shows up explicitly in the syllabus.

​

• Probability and statistics: Bayes' theorem and Hoeffding's inequality are both named, specific topics students are expected to know and apply, not just define.

​

• Calculus and optimization: derivatives, gradients, and gradient descent — the mechanism that trains almost every model students will implement later in the syllabus.

Why this section matters most: students who skip straight to "building models" without this foundation can often get a neural network to run, but they can't explain why it works, why it fails, or how to fix it when it does — and that's exactly where Round 1's theory questions are designed to catch gaps.

Classical machine learning

This is the bulk of Round 1 for most students, and it's the layer where "from scratch" understanding really gets tested.

​

• Supervised learning: linear regression, logistic regression, support vector machines (SVMs), decision trees, k-nearest neighbors, and ensemble methods. Students should understand both how to derive each method mathematically and how to apply it with scikit-learn.

​

• Unsupervised learning: k-means clustering and PCA are both explicitly on the syllabus — expect both conceptual questions and implementation tasks.

​

• Core supporting concepts: the bias–variance tradeoff, cross-validation, loss functions, and regularization. These aren't a separate topic — they show up embedded inside almost every other question, since they're how you evaluate whether any of the above methods are actually working.

​

Deep learning

The more advanced half of the syllabus, weighted more heavily in Round 2 but present in Round 1 as well.

​

• Neural network fundamentals: forward propagation, backpropagation, and the intuition for how training actually adjusts a network's weights — tested both as a derivation on paper and as working code.

​

• PyTorch implementation: building, training, and debugging real networks in PyTorch, not just describing them.

​

• Advanced architectures: transformers, natural language processing (NLP), computer vision, and generative AI fundamentals round out the more advanced material, concentrated more heavily at the Round 2 and Camp stages.

Programming environment and tools

​

• Python is the only programming language used throughout USAAIO. Fluency with syntax, data structures (lists, dicts, sets), functions, classes, and basic algorithmic thinking is a prerequisite for the ML content to make sense at all.

​

• NumPy — critically, USAAIO problems expect students to implement algorithms from scratch using NumPy, not just call a high-level library function. This is one of the biggest gaps between "took an AI course" and "is actually ready."

​

• scikit-learn for classical ML application tasks.

​

• PyTorch (2.x) for deep learning implementation.

​

• Google Colab, with theory answers typeset in Markdown and mathematical notation.

The core insight most self-study plans miss

​

USAAIO tests two skills most students only build one of — deriving an algorithm mathematically, and implementing it from scratch in code. A student who understands backpropagation only in code, without being able to derive the gradient by hand, struggles with the theory questions. A student who can prove the math but has never written a training loop in PyTorch struggles with the programming tasks.

Real preparation has to train both together, in the right order — math first, then classical ML from scratch, then deep learning — rather than jumping straight to the flashiest topic. 

Vision