Program
History and progress transfer
Session summary
Homework — optional, outside the timer
Theory of all steps of the program by blocks. Open a topic to read it.
How a session goes
A session is 10–90 minutes: a warm-up on what you've learned, new steps with short theory and problems, practice. Each problem gives two tries: 100 points on the first, 60 on the second; after the second mistake the correct answer is shown. You move on in the program if 70 % of the problems are solved.
Problems with numbers are generated anew every time a step is repeated. A fractional answer can be entered with a comma, a point or as a fraction: 0,5, 0.5, 1/2. In Practice any topic is available without a timer.
About the course
Machine learning math in practice: feature vectors, the dot product and cosine similarity, norms, the linear model y = Wx + b, MSE and cross-entropy losses, gradient descent and backpropagation on a small network, sigmoid, softmax and ReLU, logistic regression, Bayes and likelihood, quality metrics, overfitting and regularization, feature normalization and PCA. Requires linear algebra, calculus and probability.
Course program: 28 lessons
Vectors and data
- Feature vectors and the data matrix
- Dot product
- L1 and L2 norms
- Cosine similarity
- L1 and L2 norms: more
- Checkpoint: vectors and data
The linear model and training
- Linear model y = Wx + b
- MSE loss function
- Partial derivatives and the gradient
- Gradient descent
- Gradient descent: more
- Checkpoint: the linear model and training
Neural networks and backpropagation
- Activation functions: sigmoid and ReLU
- Softmax
- Chain rule
- Backpropagation
- Backpropagation: more
- Checkpoint: neural networks
Probability and classification
- Logistic regression
- Cross-entropy
- Probability, likelihood, Bayes
- Checkpoint: probability and classification
Model quality
- Metrics: accuracy, precision, recall, F1
- Overfitting and splitting the data
- Regularization
- Feature normalization
- PCA and eigenvectors
- Checkpoint: model quality