Tutorials
Tutorials will provide live-coding notebooks with learning objectives, checkpoints, exercises, verification cells, and a final interpretation section.
If an Open in Colab link cannot access GitHub, download the notebook below and use File → Upload notebook in Colab. Week 05 also needs its local assets; its Colab setup instructions fetch them from the public course website.
Week 01 — A Story of 3 and 7
Explore how representations change a machine-learning problem, using handwritten digits 3 and 7 as a running example.
Week 02 — Drawing the Line
Logistic Regression from Data to Decisions. A visual chapter using two moons and Iris to connect logits, likelihood, decision boundaries, thresholds, entropy, and softmax. Includes editable experiments and reusable figure exports; CPU only, with no dataset downloads.
Download the tutorial notebook
If direct Colab access fails, download the notebook and use File → Upload notebook in Colab. Allow 110–120 minutes for live discussion or approximately 75–90 minutes for independent study.
TODO(instructor): Review this educational draft and select figures for the Week 02 lecture.
The representative validation notebook is planned for phase 4. TODO(instructor): Confirm the supported Python environment and package policy.
SVMs and kernels self-study extends Week 02 classification and prepares for the Week 03 lab.
Week 03 — Decision trees and ensembles
The Week 03 tutorial covers ensembles: bias and variance, bootstrap, bagging, random forests and gradient boosting. Its tutorial presentation — PDF is available in Russian. The decision-trees lecture — PDF supplies the preceding material. A separate tutorial notebook is not currently published.
Week 02 SVMs and kernels self-study supports the Lab 3 competition.
Week 04 — Ode to Autograd
Use ReLU hinges and XOR to see what a hidden layer learns, then trace a forward pass and verify backpropagation against PyTorch autograd. The CPU-only notebook takes approximately 75–85 minutes with its checkpoints.
Week 05 — Convolutions: Local Patterns and Shared Weights
Follow one investigation from image tensors and fixed filters through shared trainable kernels, multiple channels, spatial controls, a small CNN, architecture reading, and a controlled MLP-versus-Conv1d experiment on MNIST-1D.
The download uses checked-in images and data under assets/; run it from a checkout of the full repository. It is not a standalone notebook bundle and does not download data at runtime.
Week 06 — Sequential Representation Learning
Complete tutorial. Follow a robot’s delayed command recall through recurrent states, gates, and causal attention. Work through calculations, train small models, and compare held-out recall across three initializations. Expand explanations and detailed ledgers as needed. Optional extensions cover coding costs and continuous-wave forecasting. Compact architecture maps trace the wave models from observed samples through learned features to a scalar output, then show recursive feedback. Controlled diagnostics test training budget, window length, and data count; a paired frequency-shift test separates local prediction from sustained forecasting. Full diagnostic retraining is optional.
Week 07 — Learning from evidence
Study route: Week 07 covers frequentist/Bayesian foundations and the VAE idea through Bayes’ theorem. In the existing notebook, VAE loss/ELBO calculations are forward reading for Week 10; dropout and practical uncertainty estimation are forward reading for Week 12.
Complete educational draft. Follow noisy measurements from a fitted line to a Bayesian posterior, then inspect a small dropout network. A self-contained regression derivation explains uncertainty over fixed weights and obtains the posterior from Bayes’ theorem. Read a VAE step by step: encoder heads, reparameterization, decoder, and the two loss terms. A 5–10 minute theory walkthrough connects the generative model, evidence, latent posterior, encoder approximation, ELBO, and implemented training loss. Train on a small thresholded MNIST subset, reconstruct held-out images, interpolate codes, and generate new images. Plotting helpers are folded at the beginning; working settings invite exploration. A paired KL-weight experiment and a short Naive Bayes exercise are optional. MNIST uses a cache or an explicitly enabled download; offline checks use labeled binary patterns. Instructor review of text, pacing, and environment remains pending.
Week 10 — Structure without labels
Draft for instructor review. Compare denoising, digit matching and anomaly detection in three 15–20-minute coding challenges. Optional activities cover t-SNE maps and VAE generation. Run on CPU with offline 8×8 digits, then report your observations in Moodle.