Here is a brief diary of topics discussed in classes (see also the class calendar for the AML course in 24/25 here, but the course will vary from it).
Slides and coding scripts are distributed after each lecture via the Google groups mailing-list. Refer to the home page on how to subscribe.
[22/09/2025] Introduction to the course, review introduction of machine learning and computer computer vision.
[24/09/2025] Recap of linear classification, fully connected and convolutional networks, including backpropagation and stochastic gradient descent.
[29/09/2025] Guest lecture by Prof. Bernt Schiele on inherent interpretability for deep learning in computer vision.
[1/10/2025] More on ConvNets: receptive fields, visualization and interpretability of convolutional filters.
[6/10/2025] Pytorch review of Multi-layer Perceptrons; training NN, including activation functions, data pre-processing, weight initialization, batch normalization, DropOut and data augmentation.
[8/10/2025] Hyper-parameter tuning; transfer learning; introduction of sequence modelling.
[13/10/2025] Sequence modelling with RNN
[15/10/2025] RNN gradient flow and LSTMs, including image captioning
[20/10/2025] Pytorch review of CNN's
[22/10/2025] Introduction of the course challenge
[27/10/2025] Attention, self-attention and applications to machine translation and image captioning; positional encoding
[29/10/2025] Masked- and multi-head self-attention; transformer networks architecture
[3/11/2025] Transformer networks for images; relative self-attention; BERT
[5/11/2025] Self-supervised learning: pre-text tasks for SSL
[10/11/2025] Self-supervised learning: approaches based on contrastive learning
[12/11/2025] Self-supervised learning: contrastive learning using only positives
[17/11/2025] Semantic segmentation and object detection with Fast R-CNN
[19/11/2025] Faster R-CNN and DETR
[24/11/2025] Hyperbolic deep learning
[26/11/2025] Multi-modal hierarchical learning with CLIP and HyCoCLIP
[1/12/2025] Midterm presentations
[3/12/2025] Graph Convolutional Networks