My Oxford AI Summer: The Full Syllabus

This summer I'm working through the Advanced Artificial Intelligence and Machine Learning track at Lady Margaret Hall, University of Oxford. The programme runs as three back-to-back sessions, each a self-contained three-week course convened by Dr Naeemullah Khan, an Instructional Assistant Professor in AI at KAUST and Associate Research Fellow at LMH.

Every session is built the same way: twelve numbered lecture-and-seminar blocks, backed by tutorials, skills sessions, and around forty hours of independent study a week. Below is the complete topic list for all three sessions, with the descriptions straight from the course handbooks.

Session 1: Generative AI and Deep Unsupervised Learning

29 June to 17 July 2026

This course explores how machines learn the structure of data without labels, with a heavy emphasis on modern generative modeling and self-supervised representation learning. It builds intuition from density estimation up through the current large-scale paradigms.

1. Unsupervised Learning Landscape and Density Estimation

Sets the stage for the whole course. Covers the goals of unsupervised learning, including learning structure, compression, representation, and generation. Introduces density estimation formally, discusses explicit versus implicit density models, and surveys the main model families covered in the course, including autoencoders, VAEs, GANs, normalizing flows, and diffusion models.

2. Autoencoders

Covers the encoder-bottleneck-decoder structure, reconstruction loss, and the role of the latent space in learning compressed representations of data. Discusses undercomplete and overcomplete autoencoders, denoising autoencoders, sparse autoencoders, and contractive autoencoders, while motivating the move toward probabilistic latent-variable models.

3. Variational Autoencoders (VAEs)

Introduces the probabilistic view of latent-variable modeling, where the encoder learns a distribution rather than a single point representation. Covers the derivation of the ELBO, the reparameterization trick for differentiable sampling, and the balance between reconstruction quality and KL regularization. Also discusses Beta-VAE and disentangled representation learning.

4. Representation Learning and Self-Supervised Objectives

Frames self-supervised learning as a way of capturing the structure of the data distribution through learned representations. Covers pretext tasks, contrastive predictive coding, mutual information maximization, and the relationship between self-supervised objectives and generative modeling.

5. GANs and GAN Variants

Provides a unified treatment of GANs, starting from the original minimax game between generator and discriminator. Covers the standard GAN objective, the Jensen-Shannon divergence formulation and its optimization issues, and improvements such as Wasserstein GAN and gradient penalty. Also surveys major GAN variants including conditional GANs, ProGAN, StyleGAN, BigGAN, Pix2Pix, and CycleGAN. Discusses practical issues such as mode collapse, training instability, evaluation with FID and IS, and precision-recall style metrics for generative quality.

6. Diffusion Models Foundations

Introduces diffusion models through the forward noising process and the reverse denoising process. Covers DDPM, the connection to score matching and Langevin dynamics, simplified training objectives, and DDIM for faster sampling. Explains why diffusion models have become a dominant paradigm in modern generative modeling.

7. Stable Diffusion and Latent Diffusion Models

Covers latent diffusion models and why performing diffusion in a compressed latent space is computationally more efficient than operating directly in pixel space. Discusses the VAE used for image compression, the UNet denoiser, CLIP-based text conditioning, and classifier-free guidance. Covers the end-to-end training and inference pipeline of Stable Diffusion.

8. Video Generation and World Models

Covers the extension of generative models from images to video, including the challenges of temporal consistency and coherent motion generation. Discusses video diffusion models and the conceptual foundations behind systems such as Sora. Connects video generation to world models, highlighting how predictive models of visual dynamics begin to approximate structured models of environments.

9. Transformers as Generative Models

Covers transformer-based generative modeling beyond text. Discusses autoregressive image generation through discrete tokenization approaches such as VQ-VAE with GPT-style transformers, and connects these ideas to early systems such as DALL-E. Also introduces flow matching as a modern alternative to diffusion and shows how transformer architectures can be applied to both discrete and continuous visual representations.

10. Tabular and Medical Foundation Models

Covers how foundation-model ideas extend beyond vision and language into other domains. Discusses TabPFN as a pretrained transformer-style model for tabular data and Tribe v2 as an example of a medical foundation model trained on multimodal clinical data. Highlights the broader generalization of unsupervised and self-supervised paradigms across data types.

11. Unsupervised and Self-Supervised Learning in NLP

Introduces the unsupervised learning paradigm in natural language processing, focusing on how modern language systems learn from raw text without manual labels. Covers distributional semantics, word embeddings, language modeling as self-supervision, encoder-based objectives such as masked language modeling, and autoregressive objectives used in large language models. Discusses major model families such as word2vec, GloVe, BERT, and GPT, with an emphasis on how unsupervised objectives drive representation learning at scale.

12. Modern NLP Foundation Models and Current Directions

Focuses on current topics in unsupervised and self-supervised NLP. Covers large-scale pretraining, instruction tuning as a downstream adaptation stage, multimodal language models, retrieval-augmented generation, contrastive language-image pretraining, and latent or token-based generative paradigms for text. Also discusses scaling laws, in-context learning, synthetic data generation, and the convergence of language modeling, reasoning, and world modeling in contemporary foundation models.

Session 2: Computer Vision

20 July to 7 August 2026

This course begins with the core foundations of visual learning through convolutional neural networks and works its way to the current research frontier, including self-supervised methods, world models, and large-scale vision foundation models.

1. Convolutional Neural Networks

Covers the building blocks of CNNs, including convolution operations, filters, feature maps, pooling, and stride. Discusses classic architectures such as LeNet, AlexNet, VGG, ResNet, and EfficientNet. Introduces important ideas such as receptive field, parameter sharing, hierarchical feature learning, and translation invariance.

2. Object Detection and Localization

Introduces object detection as a natural application of deep visual representations. Covers the evolution from early region-based and anchor-based detectors, such as R-CNN, Fast R-CNN, Faster R-CNN, and YOLO, to transformer-based detection approaches such as DETR and Deformable DETR. Discusses bounding box regression, classification, non-maximum suppression, set prediction, and bipartite matching loss.

3. Transformers in Vision

A compact lecture covering the transition from CNN-based vision to transformer-based vision. Discusses the motivation for self-attention in images, the computational challenges of applying attention to pixels, patch-based tokenization, positional encoding for 2D data, and the core architecture of Vision Transformers. Covers ViT, DeiT, training strategies such as distillation, and the reasons transformers scale well with data and model size.

4. Sequential Vision

Covers early sequential modeling approaches for visual data, including RNNs and LSTMs for image captioning and video understanding. Discusses encoder-decoder pipelines in which CNN features are passed to sequential decoders, and explains how temporal dependencies were modeled before the transformer era.

5. Video Understanding and Temporal Modeling

Covers methods for extending visual models from images to video. Includes 3D convolutional models such as C3D and I3D, two-stream networks combining RGB and optical flow, and transformer-based video models such as TimeSformer and Video Swin. Discusses temporal redundancy, motion modeling, and the challenges of learning from spatiotemporal data.

6. Generative Models for Vision I: Autoencoders and Variational Autoencoders

Introduces generative modeling through autoencoders and variational autoencoders. Covers encoder-decoder architectures, latent representations, reconstruction objectives, and the transition from deterministic encoding to probabilistic latent-variable modeling. Discusses the VAE loss, including reconstruction and KL divergence terms, and explains how VAEs can be used for generation, interpolation, and representation learning.

7. Generative Models for Vision II: GANs and Diffusion Models

Covers two major paradigms in modern generative vision modeling: GANs and diffusion models. Introduces the adversarial training setup of GANs, generator-discriminator dynamics, and important variants used in image synthesis. Then covers diffusion models, including the forward noising process, reverse denoising process, score-based generation, and why diffusion models have become central to modern image generation. Discusses their strengths, limitations, and growing importance in computer vision.

8. Attention Mechanisms in Vision

Introduces attention mechanisms in vision outside the full transformer framework. Covers spatial attention, channel attention, SENet, and non-local operations. Discusses how attention maps help models focus on the most relevant regions or feature channels in an image.

9. Representation Learning: Self-Supervised Learning

Introduces self-supervised learning for vision. Starts with early pretext tasks such as rotation prediction and jigsaw puzzles, then moves to modern instance discrimination methods such as MoCo and SimCLR. Also introduces masked image modeling approaches such as MAE and BEiT, highlighting how self-supervised methods learn useful visual representations without manual labels.

10. Contrastive Learning

Covers the theory behind contrastive learning objectives and the role of positive and negative pairs. Examines CLIP as a large-scale image-text contrastive model and DINO as a self-distillation-based representation learning framework. Discusses why these methods produce rich and transferable representations useful across a wide range of downstream vision tasks.

11. Advanced Self-Supervised Learning: I-JEPA and V-JEPA

Introduces the Joint Embedding Predictive Architecture paradigm. Covers I-JEPA for images, where the model predicts representations of masked regions, and V-JEPA for video, where the model predicts future latent representations rather than raw pixels. Discusses the efficiency and conceptual advantages of predictive learning in latent space.

12. World Models and Vision Foundation Models

Covers the idea of world models and their importance for building systems that can predict and reason about environments over time. Introduces the Ha and Schmidhuber formulation, the Recurrent State Space Model, and its role in DreamerV3. Connects these ideas to V-JEPA and discusses LeWorldModel as a joint-embedding predictive approach from pixels. The lecture concludes with a survey of foundation models for vision, including Grounding DINO for open-vocabulary detection, SAM-family models for promptable segmentation, Depth Anything for monocular depth estimation, SuperGlue for keypoint matching, and InternVideo as a large-scale video foundation model.

Session 3: Large Language Models and Natural Language Processing

10 August to 28 August 2026

This course examines the theoretical concepts of NLP and their current and potential future applications. It begins with the fundamentals and classical methods, builds through attention and transformers, and ends on large language models, their applications, and their limitations.

1. Introduction to NLP, Logistic Regression, Naive Bayes

Starts with the fundamentals of NLP and its importance. Covers logistic regression and Naive Bayes classifiers, and how these algorithms are applied to basic NLP tasks such as text classification and sentiment analysis.

2. N-grams and Language Models

Covers the construction and application of N-grams and language models, including the basics of statistical language models and how they predict the next word in a sequence. Lays the groundwork for more complex language modelling techniques.

3. Word Vectors

Studies word vector representations and how these vectors capture semantic meaning and relationships between words, which is crucial for many NLP tasks.

4. RNN, GRU, LSTM

Focuses on Recurrent Neural Networks (RNNs) and their variants, Gated Recurrent Units (GRUs) and Long Short-Term Memory networks (LSTMs). These models are essential for handling sequential data and understanding context in language processing.

5. Sequence to Sequence Models, Introduction to Attention

Explores sequence-to-sequence (Seq2Seq) models, which are pivotal for tasks like machine translation. Also covers the attention mechanism, explaining how it improves the performance of Seq2Seq models by focusing on relevant parts of the input sequence.

6. Detailed Overview of Different Kinds of Attention, Positional Encoding

Goes over various attention mechanisms, including self-attention and cross-attention, to gain a comprehensive understanding of attention. Also covers positional encoding, a technique used in attention models to incorporate the order of words in a sequence.

7. Transformers

Focuses on the transformer architecture, which has revolutionized NLP. Covers its components, such as multi-head attention and feed-forward neural networks, and why transformers are more efficient and effective than traditional RNN-based models.

8. Large Language Models

Examines transformer-based large language models like GPT and BERT. Discusses their architectures, training processes, and applications, highlighting how these models achieve state-of-the-art performance on various NLP benchmarks.

9. Multimodal Language Models

Explores models that integrate multiple types of data, such as text, images, and audio. Covers how these multimodal models enhance understanding and generation of language in diverse contexts.

10. Model Analysis

Studies techniques for analysing and interpreting NLP models. Looks at model probing and the gap between what we think the machine knows and what the machine actually knows.

11. Code Generation

Focuses on an application of NLP in generating code. Explores how models can be trained to understand and produce programming code, aiding in tasks like code completion and automated coding.

12. Prompt Engineering and Open Questions in NLP

Studies prompt engineering, a technique used to guide the output of language models effectively. Discusses open questions and future directions in NLP to think critically about ongoing challenges and innovations in the field.

All three sessions are part of the LMH Oxford Summer Programmes, convened by Dr Naeemullah Khan. Coursework is done in Python and PyTorch, with a recommended foundation in linear algebra, optimization, probability, and the core deep learning concepts from Goodfellow, Bengio, and Courville.