Andrew Ng's three-course ML specialization rebuilt for community college students. We derive every algorithm before we use it, implement from scratch before touching a library, and ask: optimized for what — and for whom?
Andrew Ng's Machine Learning Specialization on Coursera is the gold standard introduction — rigorous, clear, and accessible. We use it as a launchpad, then push further: every algorithm is derived, every dataset is examined for bias, and every project is connected to a real community problem.
Depth is not fixed. You choose your track at the start of each unit. All tracks share the same readings, case studies, and community discussions. The divergence is in how far into the implementation you go.
Understand what ML algorithms do, interpret model outputs, and identify failure modes. Use trained models confidently. Spot bias, evaluate benchmarks skeptically, and read ML papers without being overwhelmed by notation. Prepare for roles that use ML without building it.
Build end-to-end ML pipelines using scikit-learn, pandas, and matplotlib. Clean and analyze real datasets. Train, validate, and tune models. Understand the bias-variance tradeoff in practice. Deploy simple models as APIs. Prepare for data analyst or ML engineer roles.
Implement every algorithm from mathematical first principles using only NumPy. Derive gradient descent, backprop, the kernel trick, and EM from scratch before using any library. Understand every line of every model you deploy. Prepare for ML research and engineering roles.
Each unit starts with a community problem. Technical concepts arrive as tools to solve it — not as the point. The capstone unifies all six units into a single project that must serve a real, named need.
Problem: Can we predict rent prices for next year? Derive linear regression. Implement gradient descent. Understand why "fit the data" is never neutral — what counts as normal in housing data?
Problem: Should this job application proceed? Derive logistic regression and decision boundaries. Understand why false positive rates differ across groups — and why that matters more than overall accuracy.
Problem: Can a computer recognize handwritten rent checks? Derive perceptrons and multi-layer networks. Implement backpropagation from scratch. What features does the network learn, and why?
Problem: What patterns exist in neighborhood demographic data? Derive k-means and PCA. When we cluster people by data, what assumptions are we encoding? Whose similarities and differences are we measuring?
Required for all tracks. Audit a real deployed algorithm — in lending, hiring, healthcare, or criminal justice. Quantify disparate impact using real data. Document findings. Propose a remedy. Present to a community audience.
Build an ML system that addresses a real problem named by your community. Not a toy dataset. Not a pre-cleaned benchmark. Real data, real stakes, real constraints. Documentation must include a harm analysis.
The universal optimization engine. Derive it, implement it, understand when it fails. The math behind "learning."
Underfitting vs. overfitting. Why a model that memorizes training data fails on real people.
What you optimize is what you get. Cross-entropy, MSE, and why the choice is a value judgment.
L1, L2, dropout. Constraining complexity to prevent memorization. The math of generalization.
True positives, false positives, precision, recall. The four cells that determine who gets hurt.
When a "neutral" algorithm produces outcomes that differ systematically across protected groups. This is a technical concept, not only a political one.
Choosing what to measure is choosing what to value. The most impactful part of ML — and the most overlooked.
Test on what you didn't train on. How we actually measure generalization in practice.
The original ML Specialization — whose conceptual clarity and mathematical rigor we build directly on and depart from.
Equitable computing education. Read before implement. Student agency over assessment anxiety.
The "New Jim Code." When automation appears neutral but encodes and amplifies racial hierarchy.
Formal mathematical definitions of fairness — and why no single definition is universally correct.
Hidden racial bias in a widely used healthcare algorithm. Case study in Weeks 13–15.
The foundational textbook — free online. Our Track III references for mathematical depth.
How models targeting poor communities amplify inequality. Core reading for the bias audit unit.
The 2-minute question rule. Ungrading. Deep learning over shallow performance. Foundational to this course's assessment philosophy.