Day 3 - From neural networks to generative models
The third day bridges what you learned in Block 1 — training classifiers and regression models interactively — with a deeper understanding of how neural networks work and how they can be used to generate new content, not just predict labels.
For both days of the second bloc you have to fill out an reflection log you find here:
Docx : Reflection Log Template (Block 2)
You are encouraged to work in mixed teams!
You can download the lecture slides here:
Slides : Algorithmic Composition using Neural Networks
Note
The third day connects the interactive experiments from Block 1 to the Python code that runs underneath them, then introduces two foundational ideas in deep learning — the perceptron and the autoencoder — before moving into the music domain. Take the time to discuss with your group: you don’t have to understand every line of code, but you should be able to explain what each notebook is trying to do.
Table of contents
- H) From canvas to code — revisiting classification and regression
- I) The perceptron — where it all started
- J) Autoencoders — learning to compress and generate
- K) Working with symbolic music in Python
- L) Style imitation with Markov chains
- Submission
H) From canvas to code — revisiting classification and regression
In Block 1 you trained a neural network by clicking on a canvas. Now you will see the exact same task expressed in Python.
This notebook trains a small PyTorch network to predict musical notes from x/y coordinates — the same inputs your browser sketch used. It covers both a classification version (pick one of seven notes) and a regression version (predict a frequency in Hz).
Work through the notebook in your group. Focus on recognising the pipeline you already know: data → model definition → training loop → evaluation. The code makes explicit what ml5.js was doing for you automatically.
I) The perceptron — where it all started
Before multi-layer networks, there was a single neuron. This notebook tells the story of the perceptron: what it can do, what it famously cannot do, and why that limitation led directly to the deep networks you have been using.
Pay close attention to the XOR problem: a simple logical function that a single neuron cannot learn, no matter how long you train it. This is not a limitation of the data — it is a geometric fact. Understanding it will help you appreciate why adding more layers is not just a trick, but a necessity.
J) Autoencoders — learning to compress and generate
So far, every network you have used has been a discriminative model: given an input, predict a label or a number. An autoencoder is different — it learns to compress an input into a compact representation (the latent space) and then reconstruct it. This is your first encounter with a generative model.
The notebook trains an autoencoder on the MNIST handwritten digit dataset. After training, you can explore the latent space: points nearby in that space correspond to images that look similar. You can also interpolate — pick two points in the latent space and move smoothly between them to generate new images that blend the two originals.
K) Working with symbolic music in Python
From now on the subject is music. Before we can teach a network to generate music, we need to be able to represent music in Python. This section covers two notebooks that you can work through together.
K.1 — music21 basics
The music21 library lets you create, manipulate, and listen to symbolic music: individual notes, chords, rhythms, full scores.
Work through the notebook to learn how to create Note and Chord objects, assemble them into a Stream, and export to MIDI. You do not need to memorise the API — focus on understanding what each building block represents musically.
K.2 — Algorithmic composition
With music21 in hand, you can compose music using pure rules and randomness — no machine learning involved yet.
This notebook explores how to use probability distributions, consonance/dissonance rules, and structured randomness to generate short musical passages.
L) Style imitation with Markov chains
The day closes with your first statistical model of musical style. A Markov chain learns which notes tend to follow which other notes in a corpus — in this case, the music of Johann Sebastian Bach. It then generates new melodies by following those learned probabilities.
This model requires no gradient descent. The “training” is simply counting: how often does note B follow note A in Bach’s scores? The result is a transition matrix — a table of probabilities.
Closing reflection for Day 3
You have now seen three very different approaches to music generation: pure randomness and rules (K.2), statistical counting (L), and — coming tomorrow — neural networks. Before Day 4, think about: what would it take to generate music that genuinely sounds like a specific composer? Which of the approaches today gets closest, and why does it fall short?
Submission
Submit your work, i.e. your reflection log, here.