Machine Learning
Build models. Measure them honestly. Ship one.
Read Week One Free Download the 462-page workbookWhy Machine Learning?
Fitting a model is the easy part. Knowing whether the number it produced means anything is the job.
Most machine learning that fails in production did not fail because the algorithm was wrong. It failed because a column leaked the answer, because the metric hid the errors that mattered, or because nobody compared the model to the boring thing it replaced. This course spends as much time on measurement as on modelling, because that is where the difference between a working system and a demo actually lives.
What Makes This Course Different?
Every result on every page came out of code that was run. The output you read is the output the machine printed, including the times it disagreed with the point the lesson was making.
What This Programme Delivers
Skills You Will Build
Technical Skills
- NumPy arrays and pandas DataFrames
- scikit-learn end to end
- Regression, regularisation and leakage
- Logistic regression, kNN, naive Bayes, SVMs
- Decision trees, random forests, gradient boosting
- Clustering, PCA and anomaly detection
- Neural networks in numpy, then TensorFlow and Keras
- Text vectorisation and sequence models
- Serving a model over HTTP, with drift monitoring
Judgement
- Choosing the metric before choosing the model
- Spotting leakage before it flatters your results
- Reading a learning curve for what it actually says
- Knowing when the baseline is already good enough
- Calibrating probabilities, and why accuracy hides this
- Explaining a tuned model, and where explanation stops
Measurement First, Then Models
The difference between a course that teaches fit() and one that teaches you to trust the result.
| Aspect | This course | The usual approach |
|---|---|---|
| Results shown | Printed by code that was run | Written from what the result should be |
| Baselines | Every method compared to a simple one first | Introduced, then skipped |
| Metric choice | Decided before the model, per problem | Accuracy by default |
| Failure cases | Leakage, imbalance and drift given full weeks | A cautionary paragraph |
| Validation | Folds, calibration and per-segment error | One train/test split |
| Deployment | Packaged, served, validated, monitored | Left as an exercise |
| Capstone | Data the course has never touched | The dataset from the lessons |
16-Week Learning Roadmap
View Detailed Week-by-Week Curriculum →
Foundations
Weeks 1 to 3. Python for ML, data wrangling and profiling, and only the linear algebra, calculus and statistics that change what you do.
Supervised Learning
Weeks 4 to 6. Regression and regularisation, classification fundamentals and the threshold decision, then trees, forests and boosting.
Evaluation and Features
Weeks 7 and 8. Folds, leaks, imbalance, calibration and the segments an average hides, then turning raw columns into usable inputs.
Unsupervised Learning
Weeks 9 and 10. k-means, hierarchical, DBSCAN and mixtures, dimensionality reduction, and finding the rows that do not belong.
Deep Learning
Weeks 11 to 13. A network built in numpy with backpropagation derived, then Keras, convolutional networks, text and sequences.
Production and Capstone
Weeks 14 to 16. Spending a tuning budget properly, interpretability, packaging and serving, drift, and a project of your own.
What You Build Along The Way
A Churn Model In Week One
Environment, arrays, DataFrames and a working classifier, all inside the first week, so the rest of the course has something concrete to improve on.
scikit-learnBoosting Against A Straight Line
Gradient boosting measured against linear regression on the same data, including the cases where the straight line wins.
BaselinesAn Image Classifier, Twice
Convolution built from scratch, then a real classifier trained, then a pretrained base transferred, with the comparison run properly.
TensorFlowA Served Model With Monitoring
The model packaged, put behind HTTP, given input validation, watched for drift, and replaced safely.
FastAPIWhat The Capstone Asks Of You
Framing
Choose a problem, not a dataset
Start from a question somebody would pay to answer, then find data for it, rather than the other way round.
State the decision
Say what changes based on the model's output. If nothing changes, the project is not finished being designed.
Measurement
Beat a stated baseline
Name the simple alternative first, measure it, and show your model beating it by more than run-to-run noise.
Report the error you would rather not
Per-segment performance, calibration and the failure mode a headline metric hides.
Handover
Someone else can run it
A repository that installs, runs and reproduces your numbers on a machine that is not yours.
You can talk about it
The trade-offs, the things you tried that failed, and what you would do with another month.
Finish With
Where This Material Leads
The roles this course is written for:
Particularly suited to roles where you have to defend a model's numbers to somebody who did not build it.
Tools & Technologies
Core Tools
Also Covered
Every tool used is free and open source. Nothing on this course needs a paid licence.
Who This Is For
People Who Write Some Python
Loops and functions are enough
Analysts Moving Into Modelling
SQL and dashboards already familiar
Students In Quantitative Fields
Maths taught only where it changes the work
Assumes some Python. No prior machine learning needed. The maths is built up from scratch, and only where it matters.
Certificate of Completion
Certificate of Completion in Machine Learning
Issued by EDUSHARK TRAINING once you finish the sixteen weeks and submit your capstone. It records what you built, not a grade.
Start With Week One
A working model in the first week, then fifteen weeks making it trustworthy.