Coding-For-MBA

---

layout: default title: Machine Learning Curriculum Roadmap nav_order: 2 permalink: /ml-curriculum/

This roadmap outlines a phased journey from classic machine learning foundations to modern deep learning systems and production operations. Each phase highlights anchor lessons from the Coding for MBA series and recommends follow-on topics so you can keep advancing after Day 57.

📐 Looking for mathematical foundations? Check out the comprehensive ML Theory & Mathematics guide, which covers linear algebra, calculus, probability, neural networks, deep learning architectures, and all the mathematical concepts underlying this curriculum.

Phase 1 – Classic Machine Learning Foundations

Start here if you are following the Day 40–53 sequence.

The goal of Phase 1 is to master supervised learning workflows, evaluation techniques, and model selection before layering on deep learning.

Day	Lesson	Key takeaway
Day 40	Introduction to Machine Learning	Frame ML problems, manage the train/validate/test split, and measure performance with cross-validation.
Day 41	Supervised Learning – Regression	Fit linear models, tune regularisation, and interpret coefficients for business insights.
Day 42	Supervised Learning – Classification Part 1	Compare logistic regression and decision trees while diagnosing accuracy, precision, and recall.
Day 43	Supervised Learning – Classification Part 2	Ensemble methods, ROC curves, and thresholding strategy selection.
Day 44	Unsupervised Learning	Cluster customer segments and reduce dimensionality with PCA.
Day 45	Feature Engineering & Evaluation	Build repeatable feature pipelines and validate models with more nuanced metrics.
Day 46	Intro to Neural Networks	Understand perceptrons, activation functions, and gradient descent.
Day 47	Convolutional Neural Networks	Apply convolutional filters for image classification.
Day 48	Recurrent Neural Networks	Model sequential data with RNNs, LSTMs, and GRUs.
Day 49	Natural Language Processing	Build text classification pipelines with tokenisation and embeddings.
Day 50	MLOps	Package, persist, and monitor models for reliable deployment.
Day 51	Regularised Models	Contrast ridge, lasso, elastic net, and Poisson GLMs while measuring coefficient shrinkage.
Day 52	Ensemble Methods	Blend bagging, boosting, and stacking ensembles with calibrated probability estimates.
Day 53	Model Tuning & Feature Selection	Optimise hyperparameters with grid/Bayesian search and validate feature subsets via permutation importance and RFE.
Day 54	Probabilistic Modeling	Master Gaussian mixtures, Bayesian classifiers, EM, and HMM log-likelihoods to reason about uncertainty.
Day 55	Advanced Unsupervised Learning	Deploy DBSCAN, hierarchical clustering, t-SNE/UMAP-style embeddings, and autoencoder-driven anomaly detection.
Day 56	Time Series & Forecasting	Forecast seasonal demand with ARIMA/SARIMAX, Holt-Winters, and Prophet-style decompositions plus robust evaluation.
Day 57	Recommender Systems	Build collaborative filtering and matrix factorisation recommenders with implicit-feedback aware ranking metrics.

Recommended next steps after Phase 1

• Apply cross-validation and feature engineering to your own datasets, documenting experiment results in a reproducible format.
• Practice hyperparameter optimisation with grid search, random search, and Bayesian optimisation frameworks.
• Explore AutoML tooling (e.g., Auto-sklearn, H2O AutoML) to accelerate baseline model comparisons.

Phase 2 – Modern Deep Learning & Representation Learning

Deep learning expands the model families available in Phase 1. Focus on building intuition for architectures, transfer learning, and optimisation.

Day	Lesson	Key takeaway
Day 58	Transformers and Attention	Assemble encoder–decoder stacks, fine-tune pretrained checkpoints, and interpret attention heatmaps with deterministic demos.
Day 59	Generative Models	Compare autoencoders, VAEs, GANs, and diffusion denoisers while monitoring reconstruction loss curves on synthetic data.
Day 60	Graph and Geometric Learning	Prototype GraphSAGE and GAT-style message passing networks for toy node classification tasks with interpretable attention scores.
Day 61	Reinforcement and Offline Learning	Explore policy/value methods, contextual bandits, and offline evaluation baselines that converge to reproducible reward thresholds.

• Refresh Python packages for GPU acceleration (PyTorch or TensorFlow) and practice training models on cloud notebooks.
• Study convolutional network variants (ResNet, EfficientNet) and fine-tune pretrained models for domain-specific images.
• Learn about transformer architectures for text, vision, and multimodal data, starting with embeddings and attention mechanisms.
• Experiment with self-supervised learning and contrastive objectives to pre-train representations on unlabelled data.

Phase 3 – Responsible AI & Model Governance

Ensure that your models meet ethical, legal, and organisational standards before moving them into production.

• Perform bias and fairness audits using disparate impact, equal opportunity, and calibration diagnostics.
• Implement explainability techniques (feature importance, SHAP, LIME) to communicate model decisions to stakeholders.
• Establish data governance practices: data lineage, versioning, and documentation (model cards, data sheets).
• Learn the regulatory landscape (GDPR, CCPA, sector-specific rules) and how they influence data collection and model deployment.

Phase 4 – MLOps, Monitoring, and Lifecycle Management

Turn prototypes into production systems by investing in reliable infrastructure and collaboration workflows.

Day	Lesson	Key takeaway
Day 65	MLOps Pipelines and CI/CD	Orchestrate feature stores, registries, and GitHub Actions workflows with DAG-style automation.
Day 66	Model Deployment and Serving	Compare REST, gRPC, batch, streaming, and edge serving patterns with load-tested adapters.
Day 67	Model Monitoring and Reliability	Detect data drift, evaluate canaries, and export observability metrics for retraining triggers.

• Industrialise feature pipelines with tools like Feature Store platforms or workflow orchestrators (Airflow, Prefect).
• Automate training and evaluation with CI/CD pipelines, integrating unit tests, data quality checks, and model validation gates.
• Deploy models to scalable serving infrastructure (REST endpoints, batch scoring, or streaming) with containerisation and serverless platforms.
• Instrument monitoring for data drift, model drift, and performance regressions; integrate alerting and human-in-the-loop retraining.
• Track experiments, artefacts, and lineage with MLflow, Weights & Biases, or similar platforms, ensuring reproducibility across teams.

Putting it all together

Progressing through these phases transforms the Day 40–53 lessons into a comprehensive ML competency path. Loop back to earlier phases whenever you encounter new domains or stakeholders—revisiting the fundamentals will keep each new system grounded in sound methodology.

This site is open source. Improve this page.