Loading…
Loading…
Loading…
Build real machine learning systems from first principles — then discover why your first working model was lying to you, and fix it.
Haithem Mihoubi
المدرّس
دورة مجانية
سجّل فوراً — كل الدروس مشمولة
يتضمن هذا الدورة:
Most machine learning courses teach you a sequence of algorithms. You finish knowing what a random forest is, and you still cannot ship a model anyone should trust.
This course is built around a single arc. In Part 2 you build a complete, working house-price model end to end — before you have studied a single algorithm in depth. It scores well. You will be pleased with it.
In Part 4 you find out it was reading the answer off a column it should never have seen, and that everything you concluded from its score was worthless.
That experience — discovering data leakage in a model you built yourself and were proud of — is the thing that separates people who can be trusted with a model in production from people who can call .fit(). It is systematically under-taught: the ML education literature specifically flags leakage as a topic students learn as a rule ("split before you scale") without ever learning the reason, so they misapply it the moment the situation changes shape.
Two datasets follow you through all 45 hours. You meet them in Part 0 and you are still improving the same pipeline in Part 9. Nothing in this course is a self-contained lesson you can forget.
homes — predicting sale price. Mixed numeric and categorical, genuinely missing values, one deliberately planted leak.churn — predicting cancellation. Class-imbalanced, time-ordered, with a subtle temporal leak of its own.You implement mean squared error, a gradient descent step, the sigmoid, Gini impurity, and backpropagation by hand in pure Python before you are allowed to import the library version. Twenty lines of arithmetic converts .fit() from magic into mechanism, permanently.
Then you use the real libraries the way practitioners do, and every lesson ends with a lab that checks what your code outputs, not whether you can recognise a definition in a multiple-choice list.
-67%مميزGo from first principles to running a production ML platform: data versioning, feature stores, CI/CD/CT, Kubernetes serving, monitoring, drift detection, security, cost, LLMOps, and incident response.
Machine Learning: Zero to Hero
دورة مجانية
What you will build, and the honest prerequisites
Your environment and the two datasets that follow you
Rules vs. examples: what learning from data actually means
Supervised, unsupervised, and what a label really costs
Framing the problem: what are you predicting, and for whom?
First look at the data — and the split you do before anything else
Where data comes from, and why it is always dirty
Missing values: the mechanism dictates the fix
Categorical data: one-hot, ordinal, target encoding and cardinality
Overfitting and underfitting, seen not defined
The bias-variance decomposition, with a worked example
Cross-validation: k-fold, stratified, grouped, time-series
Multiple linear regression and the normal equation
Regularisation: ridge, lasso, and when shrinkage buys nothing
Decision trees: splitting, impurity, and why they memorise
Bagging and random forests: variance reduction, measured
k-means, and choosing k when there is no right answer
Density-based clustering, and where k-means structurally fails
From logistic regression to a network
Activations, width and depth
Backpropagation, worked by hand
From artefact to service
Training/serving skew: the leak that appears after deployment
Drift: the measured cost of a world that changed
لا توجد مراجعات بعد. كن الأول!
سجّل في الدورة للمشاركة في النقاش.
لا تعليقات بعد. ابدأ النقاش!
-67%مميزGo from the agentic loop to governed, production multi-agent systems: tool use, planning, memory, MCP, agentic RAG, multi-agent orchestration, safety, sandboxing, observability, and enterprise deployment.
-67%مميزGo from first principles to running production NLP systems: tokenization, embeddings, transformers, fine-tuning, RAG, prompt engineering, alignment, evaluation, safety, and enterprise-grade serving.
مميزA 180-hour beginner-to-expert AI engineering program covering software, data, ML, deep learning, vision, NLP, LLMs, RAG, agents, MCP, MLOps, cloud, security, architecture, and AI SaaS.
What 'best fit' means — building the loss function by hand
Gradient descent: the loop underneath .fit()
A baseline you must beat
Minimal preprocessing: impute, encode, and keep it inside the pipeline
Train, evaluate, iterate
Ship it: save the model and write the prediction function
Numerical data: scaling, skew and outliers
Feature engineering: what the model cannot derive for itself
Class imbalance: why 89% accuracy catches 2% of churners
Pipelines: making the leak structurally impossible
Data documentation: what you must record for anyone to trust this
Data leakage part 1: preprocessing before the split
Data leakage part 2: target leakage in your Part 2 pipeline
Choosing a metric that matches the decision
Confidence in a number: repeated CV and bootstrap tests
Logistic regression: from odds to a decision boundary
Thresholds, ROC and precision-recall curves
k-nearest neighbours and the curse of dimensionality
Naive Bayes and why 'naive' still works
Support vector machines: margins, kernels, and an honest result
Hyperparameter search, and the nested loop most people skip
Boosting: fitting the residual, and knowing when to stop
Feature importance, and the metric that misleads
When to stop: the full comparison
PCA: what a principal component actually is
t-SNE and UMAP: powerful, and easy to over-read
Anomaly detection: finding the unusual without labels
Training in practice: learning rate, batches and early stopping
Head to head on churn: does the network win?
When a neural network is the wrong tool
Monitoring, retraining triggers and rollback
Fairness: the subgroup measurement your aggregate metric hides
Model cards, limitations, and refusing to ship