ICML study guide
ICML · Work through your exam, one topic at a time.
The ICML (International Certified Machine Learning) credential is a vendor-neutral certification for practitioners who build, evaluate, and operate machine learning systems in production settings. It suits data scientists, ML engineers, and technical leads who want an independent signal that their knowledge spans the full lifecycle rather than a single library or cloud platform. The ICML study guide has 8 sections and about 1,881 words.
01Mathematics and Statistics Foundations
Expect questions that test whether you can pick the right tool and reason about its assumptions, not just compute. Core topics: probability (conditional probability, Bayes' theorem, independence vs.
236 words, about 1 min02Data Preparation and Feature Engineering
Questions probe whether you understand why each preprocessing step exists and when it can leak information. Cover: missing-data mechanisms (MCAR, MAR, MNAR) and imputation strategies; scaling (standardization, min-max, robust scaling) and which models require it; encoding categoricals (one-hot, ordinal, target/label encoding and its leakage risk); feature transformations (log, Box-Cox, binning); handling outliers; class imbalance (resampling, class weights, SMOTE); and the cardinal rule that all fitted transformations must be learned on training data only and applied to validation/test.
270 words, about 1 min03Supervised Learning
Expect coverage of the bias-variance tradeoff, regularization (L1 sparsity, L2 shrinkage, elastic net), linear and logistic regression (link functions, odds, log-loss), decision trees and their instability, ensembles (bagging reduces variance, boosting reduces bias), random forests, gradient boosting, SVMs (margins, kernels, C), k-NN (curse of dimensionality, distance choice), and naive Bayes (conditional independence assumption). Understand loss functions and how model capacity interacts with data size.
238 words, about 1 min04Unsupervised Learning
Topics include k-means (objective, initialization sensitivity, choosing k via elbow or silhouette), hierarchical clustering (linkage criteria), DBSCAN (density, eps, minPts, noise handling), Gaussian mixture models and the EM algorithm (soft assignments, covariance structure), PCA (variance maximization, orthogonal components, explained variance, standardization prerequisite), t-SNE and UMAP (visualization only, no global distances), and anomaly detection. Also expect dimensionality-reduction tradeoffs and evaluation without labels (silhouette, Davies-Bouldin).
241 words, about 1 min05Model Evaluation and Validation
Master the confusion matrix and derived metrics (accuracy, precision, recall, F1, specificity, ROC-AUC, PR-AUC), when PR curves beat ROC (imbalanced data), cross-validation schemes (k-fold, stratified, grouped, time-series split), the bias-variance decomposition of error, learning curves, and statistical comparison of models (paired tests, confidence intervals). Understand calibration, threshold selection tied to business costs, and the difference between validation and test sets.
236 words, about 1 min06Deep Learning and Neural Networks
Expect backpropagation and the chain rule, activation functions (ReLU and its variants, sigmoid, tanh, softmax), loss functions (cross-entropy, MSE), optimizers (SGD, momentum, Adam), regularization (dropout, weight decay, early stopping, batch/layer normalization), initialization, vanishing/exploding gradients, and architecture families: CNNs (convolution, pooling, receptive field), RNNs/LSTMs, and Transformers (self-attention, positional encoding). Understand overfitting signals and hyperparameter effects.
222 words, about 1 min07ML Engineering, Deployment, and MLOps
Questions cover the ML lifecycle: experiment tracking, reproducibility (seeds, versioned data and code), feature stores, model registries, CI/CD for models, containerization, serving patterns (batch, real-time, streaming), latency/throughput tradeoffs, model compression (quantization, pruning, distillation), and monitoring for data drift, concept drift, and performance decay. Also expect A/B testing, shadow deployment, canary rollouts, and rollback strategies.
214 words, about 1 min08Responsible AI, Ethics, and Governance
Expect fairness definitions and their incompatibility (demographic parity, equalized odds, predictive parity cannot all hold simultaneously except in degenerate cases), bias sources (historical, sampling, label, measurement), transparency and explainability (feature importance, SHAP, LIME, inherently interpretable models), privacy (anonymization limits, differential privacy, k-anonymity), consent and data minimization, accountability, and regulatory frameworks (risk-based governance, documentation such as model cards and datasheets). Understand human-in-the-loop and contestability.
224 words, about 1 min