What the olympiad covers.

1. 01

   Foundational Skills & Classical Machine Learning

   Python and its data libraries, the classic models that learn from examples, and how to tell whether a model is any good.

   25 topics

   Programming Fundamentals

   • Python Basics (Loops, Functions, etc.)
   • NumPy and Pandas for Data Handling
   • Matplotlib and Seaborn for Visualization
   • Scikit-learn for ML
   • PyTorch Basics
   • Tensor (Multi-dimensional Array) Manipulation
   • Training Models on CPU and GPU

   Supervised Learning

   • Linear Regression
   • Logistic Regression
   • L1 & L2 Regularization
   • K-Nearest Neighbors (K-NN)
   • Decision Trees
   • Model Ensembles (Gradient Boosting, Bagging, Random Forest)
   • Support Vector Machines (SVM)

   Unsupervised Learning

   • K-Means Clustering
   • Principal Component Analysis (PCA)
   • t-SNE, UMAP
   • DBSCAN, Hierarchical & Spectral Clustering

   Data Science Fundamentals

   • Model Evaluation Metrics (Accuracy, Precision, Recall, F1-Score, etc.)
   • Underfitting, Overfitting
   • Hyperparameter Tuning
   • Cross-Validation
   • Confusion Matrix and ROC Curves
   • Feature Engineering Feature Engineering involves transforming raw, potentially high-dimensional data, categorical data, time series, or ragged data into a compact set of informative features. Techniques involve sliding windows, pooling operations, one-hot encoding, statistical moment-based features (average, standard deviation), PCA and neural-network-based embeddings.
   • Data Processing Data Processing concerns the handling of missing data and irregular data, including in sequence modeling settings. Techniques involve basic imputation strategies (mean/median/forward-fill for sequences) and padding for variable-length sequences. Covered here are also normalization and standardization techniques, train/validation/test splitting strategies, basic data augmentation (flipping, cropping, noise addition), tokenization and vocabulary building for text and audio and patching for images.

2. 02

   Neural Networks & Deep Learning

   How a neural network learns, the architectures behind today’s AI, and how to train them well.

   18 topics

   Neural Networks

   • Perceptron Basics
   • Gradient Descent
   • Backpropagation
   • Activation Functions (ReLU, Sigmoid, Tanh)
   • Loss Functions (MSE, MAE, Cross Entropy, etc.)

   Deep Learning

   • Multi-Layer Perceptrons (MLP)
   • Data Embeddings (text, image, audio)
   • Pooling Techniques (Max, Average)
   • Attention Mechanism
   • Transformers (theory needed only for text and image)
   • Autoencoders
   • SGD, Mini-Batch Gradient Descent
   • Momentum Methods (Adam, AdamW)
   • Convergence and Learning Rates
   • Regularization: Dropout, Early Stopping, Weight Decay
   • Weight Initialization
   • Batch Normalization
   • Model Finetuning (full and parameter-efficient)

3. 03

   Computer Vision

   Models that see, from classifying images to finding what is in them and generating new ones.

   10 topics

   Fundamentals

   • Convolutional Layers
   • Image Classification
   • Object Detection (YOLO, SSD, DETR)
   • Image Segmentation (U-Net)
   • Pre-trained Vision Encoders (e.g. ResNet)
   • Image Augmentation
   • Generating Images with GANs
   • Self-Supervised Learning for Vision
   • Vision-text encoders (e.g. CLIP)
   • Diffusion Models

4. 04

   Natural Language Processing & Audio

   Text and sound through pre-trained models, from classifying text to language models and speech.

   7 topics

   NLP

   • Text Classification
   • Pre-trained Text Encoders (e.g. BERT)
   • Language Modeling
   • Encoder-Decoder Models (e.g. for Machine Translation or Vision-Language Modeling)
   • Pre-trained Language Models (open-source and API-based ones)

   Audio Processing

   • Pre-trained Audio Encoders: HuBERT
   • Audio Models: Qwen-Audio, Whisper, Voxtral

The data can be text, tabular, image, audio, video, and time-series, and should be processed with the methods above. Topics not listed are considered excluded from the syllabus. Questions can be directed to the International Scientific Committee at isc@ioai-official.org.