Week 1 · Day 2 · Tuesday 13 October

Foundations of Modern Computer Vision

Learning from unlabelled images: pretext tasks, contrastive learning and self-distillation with SimCLR, BYOL, DINO, DINOv2 and iBOT, then masked autoencoders and JEPA latent prediction.

Topic Description & Objectives Resource
Self-Supervised Learning Why learn without labels, the cognitive principles behind it, a taxonomy of pretext tasks, reconstructive tasks, and instance discrimination Open
Contrastive Learning Methods SimCLR, BYOL, DINO and DINOv2, iBOT, and weakly supervised pretraining at scale Open
Advanced SSL and JEPA Models Masked autoencoders and representation collapse, I-JEPA, how SSL encoders are evaluated, and V-JEPA and V-JEPA 2 for video, planning and language alignment Open
Lab 1: Self-Supervised Pretraining by Rotation Exercise: with only a small labelled subset of CIFAR-10, compare rotation-prediction pretraining against transfer learning and training from scratch Notebook Colab
Lab 2: SimCLR Contrastive pretraining on STL-10 from two augmented views of each image, then a linear probe against a random-encoder baseline Notebook Colab
Lab 3: BYOL Self-supervised learning without negative pairs: an online network and a slowly updated target network on STL-10, evaluated with a linear probe Notebook Colab
Lab 4: I-JEPA Predict the features of masked image regions instead of their pixels, with an EMA target encoder, then linear-probe the features on STL-10 against an autoencoder's Notebook Colab
Optional: Introduction to Deep Unsupervised Learning Supervised versus unsupervised learning, clustering with k-means, hierarchical clustering and DBSCAN, dimensionality reduction with PCA, t-SNE and UMAP, and anomaly detection Open
Optional: Custom Datasets and Image Augmentations Load a folder of raw images into a PyTorch Dataset, turn them into tensors, and apply torchvision transforms and augmentations Notebook Colab
Optional: Dataset Class Summary A reference Dataset class for each common layout: class folders, pre-split or unsplit data, CSV labels, and segmentation masks Notebook Colab
Optional: Working with Different Dataset Structures Exercise: inspect real datasets stored in different layouts, write a Dataset class for each, and visualise samples Notebook Colab
Optional: Working with Different Dataset Structures (Solution) Worked solution to the dataset structures exercise Notebook Colab