‹ All tracks

Machine Learning

Learning from data in plain Python: regression and gradient descent, k-NN, metrics, logistic regression, Naive Bayes, decision trees, k-means and the perceptron.

• Online Python compiler
• Python code visualizer
• Machine Learning exercises with solutions

1. 1

   What machine learning is

   Learning rules from examples instead of writing them: supervised, unsupervised and reinforcement learning, and the words features, labels, model, training and prediction.

2. 2

   Data, statistics and features

   Datasets as rows and columns, mean, median and standard deviation, scaling features with min-max and z-scores, one-hot encoding, and missing values.

3. 3

   Linear regression

   Predict a number with a straight line: residuals, mean squared error, the least-squares formulas for the best line, more features, and reading the result.

4. 4

   Gradient descent

   Train a model by walking downhill on its loss: the gradient, the update rule, the learning rate, and batch, stochastic and mini-batch gradient descent.

5. 5

   Training, testing and overfitting

   Judge a model on data it has never seen: train and test sets, overfitting and underfitting, the bias-variance trade-off, validation sets and cross-validation.

6. 6

   k-nearest neighbours

   Classify a new example by a vote of the k most similar training examples: distances, choosing k, why scaling matters, and the cost of a lazy learner.

7. 7

   Measuring a classifier

   Accuracy and its traps, the confusion matrix, precision, recall and the F1 score, and which one matters for spam filters and medical tests.

8. 8

   Logistic regression

   Classification with probabilities: the sigmoid function, the decision boundary, the log loss, and training by gradient descent.

9. 9

   Bayes' rule and Naive Bayes

   Bayes' rule and the base-rate trap of medical tests, then a Naive Bayes spam filter built from word counts, with Laplace smoothing and log probabilities.

10. 10

    Decision trees

    A tree of yes/no questions: entropy, information gain and the Gini index for choosing each question, overfitting and depth limits, and random forests.

11. 11

    Clustering with k-means

    Unsupervised learning: group unlabelled points by alternating two steps, assign and update; why the start matters; and choosing k with the elbow method.

12. 12

    Perceptrons and neural networks

    An artificial neuron, the perceptron learning rule, why one neuron cannot learn XOR, hidden layers, activation functions, backpropagation, and where deep learning goes next.

13. 13

    Machine learning in practice

    The steps of a real ML project, the libraries professionals use, data leakage, and the questions of fairness, privacy and accountability every model raises.