Syllabus

Updated: 28 September 2026.

1. Foundational Mathematics and Programming

  • Linear algebra
  • Calculus
  • Probability
  • Statistics and estimation
  • Numerical computation
  • Discrete mathematics and graphs
  • Python programming
  • NumPy and array manipulation
  • Pandas and structured data
  • Visualization

2. Artificial Intelligence Foundations

Topic Expected competency
Intelligent agents Agents, environments, observations, actions, goals, utility, rationality, and the perception-action loop.
Search Uninformed and heuristic search; completeness, optimality, computational complexity, and the design and evaluation of heuristics.
Adversarial search Reason about competing actions, evaluate game states, reduce unnecessary exploration, and balance search depth against compute cost.
Constraint satisfaction Represent variables, domains, and constraints; find feasible assignments and reduce the search space.
Planning State-space planning, action models, goals, planning graphs, and the relation between planning and search.
Knowledge representation and logic Propositional and first-order logic, inference, rule systems, ontologies, and limitations of symbolic representations.
Probabilistic reasoning Represent dependencies between random variables, perform exact or approximate inference, and update beliefs from sequential observations.
Decision theory Expected utility, risk, value of information, and decisions under uncertainty.
Multi-agent systems and game theory Cooperation, competition, equilibria, basic mechanism design, and decentralized decision making.

3. Machine Learning Foundations and Practice

Topic Expected competency
Prediction tasks Model continuous targets and class probabilities; understand feature transformations, decision boundaries, residuals, model complexity, and interpretation.
Data preparation and feature engineering Handle missing values, scale inputs, encode categorical data, construct informative features, and build preprocessing and model pipelines.
Information theory and learning objectives Measure uncertainty, information shared between variables, and differences between distributions; relate these quantities to learning objectives.
Statistical estimation and uncertainty Estimate model parameters and predictive uncertainty, incorporate prior knowledge, and account for uncertainty when comparing multiple hypotheses.
Optimization Use zeroth-, first-, and second-order information to update parameters, choose how much data to use per update, handle constraints, and diagnose convergence and numerical stability.
Regularization Control model complexity and overfitting through training constraints, data variation, and validation-based model selection.
Hyperparameter optimization Explore hyperparameter settings with efficient search strategies, allocate search budgets, and decide when to stop trials using validation results.
Validation and model evaluation Separate training, validation, and test data; respect group structure, prevent leakage, isolate the effects of changes, and reproduce experiments. Select metrics for classification, ranking, and regression; assess calibration and subgroup errors.
Class imbalance and cost-sensitive learning Account for unequal class frequencies and error costs when preparing data, defining objectives, and selecting decision thresholds.
Semi-supervised and weakly supervised learning Learn from combinations of labeled, unlabeled, and imperfectly labeled data; assess the reliability of supervision and resulting predictions.
Domain shift and generalization Covariate shift, label shift, domain adaptation, invariance, and robust validation.
Efficient data pipelines Caching, batching, parallel loading, memory mapping, sparse data, and profiling time and memory use.

4. Classical Statistical Learning

Topic Expected competency
Similarity-based prediction Predict from similar examples; choose distance measures, account for feature scale, and balance local detail against computational cost.
Tree-based models and ensembles Partition feature spaces, control model complexity, handle missing values, and combine predictions from multiple models; assess interpretability and robustness.
Margin-based and kernel methods Understand separation margins, nonlinear similarity functions, decision boundaries, and scaling considerations.
Probabilistic classifiers Model class probabilities and feature distributions, assess assumptions about dependencies, and interpret predictive uncertainty.
Clustering Group similar examples, choose similarity measures and grouping assumptions, handle noise, and assess cluster quality.
Dimensionality reduction Select or transform features into compact representations; assess retained information and the limits of low-dimensional visualizations.
Anomaly and novelty detection Model expected behavior, score departures from it, and select detection thresholds using suitable validation data.

5. Neural Networks and Deep Learning

Topic Expected competency
Feedforward networks Layered computation, nonlinear representations, approximation capability, and network capacity.
Activations and losses Choose activation functions and training objectives suited to the target and output constraints; understand their effects on gradients and prediction errors.
Neural network training and diagnosis Compute gradients through computational graphs, update parameters, and control update sizes; vary training difficulty or sequence length, and diagnose instability and forgetting using learning curves, gradient behavior, and controlled experiments.
PyTorch and CPU/GPU execution Tensor manipulation, datasets and data loaders, modules, losses, optimizers, training loops, evaluation mode, checkpointing, random seeds, device placement, batch sizes, mixed precision, and memory management.
Initialization and normalization Choose initial parameters and normalize inputs or internal representations; understand their effects on gradient flow and training stability.
Embeddings and pooling Learn representations for text, images, audio, and categorical data; aggregate variable-sized inputs into useful representations.
Convolutional networks Local feature extraction, shared parameters, spatial resolution, receptive fields, and applications to images and time series.
Recurrent sequence models Maintain and update hidden state, process variable-length sequences, and understand how gradients behave over long dependencies.
Attention and transformers Attention-based interactions between inputs, position information, visibility constraints, information flow between layers, and training stability.
Reconstruction-based representation learning Learn compact representations by reconstructing inputs; assess reconstruction quality, robustness to corrupted inputs, and generative use.
Transfer learning and fine-tuning Reuse pretrained representations, prepare adaptation data, and choose which parameters to update under memory and compute constraints; assess transfer quality and loss of previously learned capabilities.
Self-supervised representation learning Construct supervision from data, learn useful similarities and invariances, and evaluate representations on downstream tasks.

6. Reinforcement Learning and Sequential Decision Making

Topic Expected competency
Multi-armed bandits Balance exploration and exploitation, use uncertainty to guide action selection, and assess cumulative performance relative to the best action.
Markov decision processes and value functions States, actions, transitions, rewards, policies, returns, discounting, and episodic or continuing tasks; relate state and action values to future rewards.
Planning with known dynamics Evaluate and improve policies using known transitions and rewards; understand convergence and computational trade-offs.
Value estimation from experience Estimate future returns from complete or partial trajectories, combine observed rewards with current value estimates, and understand estimation bias and variance.
Learning control policies Improve action selection from experience, distinguish data collected by the current policy from data collected by other policies, and manage exploration.
Function approximation and deep RL Approximate values or policies with neural networks, reuse collected experience, and diagnose instability caused by changing targets and correlated data.
Policy optimization Optimize expected return directly, use value estimates to guide policy updates, reduce update variance, and balance improvement against policy stability.
Reward design Define rewards for goals and constraints, handle sparse feedback, and recognize unintended incentives.
Imitation, offline, and model-based RL Learn from demonstrations or fixed datasets, account for limited data coverage, learn environment dynamics, and plan with imperfect models.
RL evaluation and reproducibility Evaluate across multiple random seeds and episodes; compare return distributions, sample efficiency, and training stability.
Multi-agent reinforcement learning Cooperative and competitive settings, non-stationarity, centralized training, and decentralized execution.

7. Computer Vision

Topic Expected competency
Image processing and augmentation Pixels, channels, color spaces, spatial resolution, input scaling, and controlled image transformations that preserve valid labels.
Image classification Train and adapt classifiers to assign image-level labels using pretrained image representations.
Object detection Localize and classify objects, represent bounding boxes, handle duplicate predictions, and evaluate localization and recognition quality.
Image segmentation Assign labels to pixels and distinguish object instances; choose suitable training objectives and evaluate region overlap and boundary quality.
Generative vision Learn image distributions, condition generation on inputs, produce samples, and understand limitations of generation quality and evaluation.
Video and temporal vision Frame sampling, temporal pooling, motion features, sequence models, and video classification.

8. Natural Language Processing, Speech, and Audio

Topic Expected competency
Text preprocessing and tokenization Sentence, word, and subword tokenization, vocabularies, padding, masking, and sequence-length handling.
Text representations and encoders Frequency-based and learned text representations, contextual information, similarity measures, sequence aggregation, and adaptation to downstream tasks.
Text classification and sequence labeling Assign labels to documents or individual tokens; handle label dependencies and align token predictions with the original text.
Language modeling and sequence-to-sequence learning Predict text from preceding or surrounding context, map input sequences to output sequences, assess predictive likelihood, and apply models to translation and summarization.
Retrieval and semantic search Dense and sparse retrieval, embeddings, similarity, reranking, and retrieval evaluation.
Audio and signal representations Waveforms, sampling rate, resampling, framing, time-frequency representations, and normalization.
Speech and audio models Use provided models to recognize, transcribe, or interpret speech and other audio; prepare inputs and evaluate outputs for the task.

9. Foundation Models and Large Language Models

Topic Expected competency
Prompting and context management Design instructions and examples, organize complex tasks, select relevant context, handle context limits, and evaluate prompt effectiveness.
Structured generation and decoding Choose outputs from predicted distributions, balance diversity and reliability, enforce output constraints, validate and repair outputs, and define stopping criteria.
Preference and reinforcement-based adaptation Learn from preferences and reward signals, optimize model behavior, and assess trade-offs between task performance and adherence to desired behavior.
Retrieval-augmented generation Select and organize retrieved evidence for generation, ground responses in that evidence, and evaluate factual support and citation accuracy.
Tool use and agents Function/tool calling, planning loops, memory, state tracking, environment feedback, and rule-constrained actions.
LLM evaluation Assess output validity, factual accuracy, hallucination, robustness to prompt variations, human preferences, and contamination risks.
Inference efficiency Reduce computation and memory use, process requests efficiently, reuse intermediate results, and balance sequence length, model size, latency, and output quality.
Multimodal foundation models Joint text-image-audio inputs, modality encoders, fusion, prompting, and cross-modal evaluation; image-text similarity, retrieval, and classification using text descriptions.
LLM safety and security Prompt injection, jailbreaks, data leakage, unsafe outputs, tool misuse, and defensive validation.

10. Learning from Time Series, Sequences, and Sensor Data

Topic Expected competency
Sampling and preprocessing Sampling frequency, windows, overlap, resampling, synchronization, variable-length sequences, event segmentation, filtering, detrending, outliers, and leakage-safe transformations.
Statistical and frequency-domain features Summarize distributions, temporal dependence, periodicity, and signal energy to construct informative features.
Classical forecasting Model trends, seasonality, and dependence on past observations; establish baseline forecasts and assess modeling assumptions.
Multi-step forecasting Use multivariate input sequences to predict individual future values or output sequences; predict multiple steps jointly or feed predictions into later steps, and assess training and inference differences, error accumulation, and horizon selection.
Temporal validation and metrics Respect chronology and event boundaries, evaluate across forecast periods, choose error measures suited to scale and horizon, and prevent leakage.
Anomaly and change-point detection Identify unusual observations in temporal context and changes in behavior over time; distinguish isolated anomalies from sustained shifts.
Sequence similarity and alignment Compare and align sequences while accounting for shifts in timing, phase, and speed; assess the suitability of similarity measures.
Learning from sensor and IMU data Accelerometer/gyroscope data, axes, orientation effects, sensor fusion, activity recognition, and biometric verification.

11. Contest Tasks

What to Expect

  • Tasks involve implementing and adapting solutions, training or using models, producing predictions, or choosing actions in interactive environments within the time and compute budget.
  • A task may combine tabular data, images, text, audio, video, time series, sensor streams, or graphs.

What Not to Expect

  • Solving a task simply by applying an off-the-shelf model or method unchanged.
  • Questions that depend on memorizing obscure library APIs.

These resources support preparation. The ICAIC topic tables above define the examinable scope; the recommended reading does not add requirements.

Resource Subject
Artificial Intelligence: A Modern Approach — Stuart Russell and Peter Norvig Classical AI.
The Elements of Statistical Learning — Trevor Hastie, Robert Tibshirani, and Jerome Friedman Statistical learning.
Reinforcement Learning: An Introduction — Richard S. Sutton and Andrew G. Barto Reinforcement learning.
Deep Learning — Ian Goodfellow, Yoshua Bengio, and Aaron Courville Deep learning.

Supplementary Resources

This site uses Just the Docs, a documentation theme for Jekyll.