Cours - Du Machine Learning aux réseaux convolutifs
A ten-hour machine learning course for third-year bachelor students: from the supervised learning pipeline through to CNNs, U-Net and variational auto-encoders.
Course material for Du Machine Learning aux réseaux convolutifs — CNN, U-Net, auto-encodeurs et VAE, ten hours taught to third-year bachelor students. The slides are in French.
The course runs over four sessions:
| Session | Hours | |
|---|---|---|
| 1 | La méthode ML — data, loss, optimisation, evaluation | 2 h |
| 2 | Les maths, et le gradient — tensors, matrices, backpropagation, convolution | 2 h |
| 3 | CNN et images — learned filters, feature maps, training a CNN | 3 h |
| 4 | CNN avancé, U-Net, espace latent — ResNet, transfer, segmentation, compression | 3 h |
The four sessions are built around a single idea: you define a model with parameters, a cost function that measures its error, and you descend the gradient. CNN, U-Net and VAE are then different choices of model and cost.
The lab sheets are in preparation and will be added here.