This page shows a history of previous sessions in the reading group.
| Date | Topic | Room | Lead |
|---|---|---|---|
| 20/03/23 | Introduction to word embeddings and language modelling (Slides) | David Blackwell | Fede Nanni |
| 03/04/23 | Deep Learning Basics (Slides) | David Blackwell | Phil Swatton, Jack Roberts |
| 17/04/23 | Sequence-to-sequence models part I: RNNs/LSTMs (Slides) | David Blackwell | Ryan Chan |
| 03/05/23 | Sequence-to-sequence models part II: Encoder-decoder models (Slides) | David Blackwell | Ryan Chan |
| 15/05/23 | Hands-on RNN/LSTM session (Materials) | David Blackwell | Nathan Simpson, Levan Bokeria, David Llewellyn-Jones |
| 31/05/23 | Reginald overview & Attention and self-attention networks (Notebook) | David Blackwell | Evelina Gabasova, Martin Stoffel |
| 26/06/23 | Attention (continued) (Slides) & Transformer Encoder and Decoders (Slides) | David Blackwell | Martin Stoffel, Ryan Chan |
| 10/07/23 | BERT: Masked Language modelling and Pre-training (Slides) | David Blackwell | Ryan Chan |
| 24/07/23 | GPT: Pretraining Decoders (Slides) | David Blackwell | Ryan Chan |
| 07/08/23 | Vision Transformers part I (Slides) | David Blackwell | Katie Awty-Carroll, Ryan Chan |
| 21/08/23 | Vision Transformers part II (Slides) | David Blackwell | Katie Awty-Carroll |
| 18/09/23 | LoRA (+ parameter efficient fine-tuning) part I (Slides) | David Blackwell | Jack Roberts |
| 25/09/23 | LoRA (+ parameter efficient fine-tuning) part II (Notebook) | Margaret Hamilton | Jack Roberts |
| 02/10/23 | Reinforcement Learning Human Feedback (RLHF) (Slides) | David Blackwell | Eseoghene Ben-Iwhiwhu |
| 16/10/23 | Prompt Engineering (Slides) | David Blackwell | Martin Stoffel |
| 30/10/23 | Knowledge retrieval (Slides) | David Blackwell | Praveen Selvaraj |
| 06/11/23 | Discussion: Current challenges and future directions in safety evaluations for generative AI (Slides) | David Blackwell | Jonathan Bright |
| 13/11/23 | Introduction to Diffusion models (Slides) | David Blackwell | Edmund Dable-Heath |
| 20/11/23 | Research at Turing: Transformers for coding/software engineering (Slides) | Mae Jemison | Anastasiia Grishina |
| 04/12/23 | Discussion: Best Practice for Responsible Foundation Models – What Should Developers Do and How You Can Help (Slides) | Ursula Franklin | Carolyn Ashurst |
| 11/12/23 | Stable Diffusion (Slides) | David Blackwell | Edmund Dable-Heath |
| 08/01/24 | Discussion: Benchmarking AI applications on GPUs (Slides) | David Blackwell | Tomas Lazauskas, David Llewellyn-Jones |
| 15/01/24 | Retentive Networks (Slides) | David Blackwell | Ed Gunn |
| 22/01/24 | Research at Turing: Spatial Graph Patterning of Filamentous Structures | David Blackwell | Kristina Ulicna |
| 29/01/24 | Vision Transformers Need Registers (Slides) | David Blackwell | Tom Davies |
| 05/02/24 | Discussion: Existential Risk of AI? (Slides) | David Blackwell | Levan Bokeria |
| 12/02/24 | Mechanistic interpretability (Slides) | David Blackwell | Praveen Selvaraj |
| 19/02/24 | Research at Turing: Longitudinal NLP (Slides) | David Blackwell | Jenny Chim, Talia Tseriotou |
| 26/02/24 | Research at Turing: Machine translation quality estimation (Slides) | David Blackwell | Radka Jersakova, Jo Knight |
| 04/03/24 | Discussion: Expanding participatory governance for LLMs: case studies from BigCode, Aya Initiative, and Collective Intelligence Project (Slides) | David Blackwell | Jennifer Ding |
| 11/03/24 | Research at Turing: Applying Vision Transformers in Neuroscience (Slides) | David Blackwell | Bryan Li |
| 18/03/24 | Research at Turing: Not even a Chinese Room: evaluating LLMs on code simulation | David Blackwell | Emanuele La Malfa |
| 08/04/24 | Paper overviews (Slides, Slides) | Ursula Franklin | Fede Nanni, Markus Hauru, Praveen Selvaraj |
| 15/04/24 | Research at Turing: Natural Logic-based Fact Verification with LLMs | David Blackwell | Marek Strong |
| 22/04/24 | Research at Turing: Learn how to learn and distil during learning - Using meta-learning and second order optimisation to prune the model | David Blackwell | Yilei Liang |
| 29/04/24 | Invited Talk: How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions (Slides) | David Blackwell | Lorenzo Pacchiardi |
| 13/05/24 | Overview of LLM Security (Slides) | David Blackwell | Ed Chapman, Burak Hasircioglu, Ezzeldin Zaki |
| 20/05/24 | KAN: Kolmogorov-Arnold Networks | Ursula Franklin | Andrew Duncan |
| 04/06/24 | Invited Talk: Are we ready for attacks on machine learning? | Enigma (2.30pm) | Nicholas Carlini |
| 01/07/24 | A perspective on the fundamentals of transformers (Slides) | Ursula Franklin | Ed Gunn |
| 08/07/24 | Invited Talk: Equally Safe Online? A participatory approach to tackling Gender-Based Violence (Slides) | David Blackwell | Gavin Abercrombie |
| 15/07/24 | Invited Talk: Open science projects for open-source models and transparent open datasets | Cipher | Christopher Klamm |
| 22/07/24 | Invited Talk: Designing a Value-driven GAI Framework for Social Good: Embedding Social Good Values into GAI Models (Slides) | Ursula Franklin | Victor OK Li, Jacqueline CK Lam and Jon Crowcroft |
| 05/08/24 | Invited Talk: The growth of parallelism in machine learning inference (Slides) | Ursula Franklin | Tim Harris (Microsoft) |
| 12/08/24 | Llama 3.1 Report Overview (Slides) | Ursula Franklin | Edwin Brown, Ryan Chan |
| 19/08/24 | Overview of Knowledge Graphs (Slides) | David Blackwell | Navdeep Kaur |
| 28/08/24 | Mixture of Experts (Slides) | Jack Good | Angus R Williams |
| 03/09/24 | Invited Talk: Sociotechnical Safety Evaluation of AI systems (Slides) | Enigma | Laura Weidinger |
| 09/09/24 | Invited Talk: On the Brittleness of Prompts in LLMs (Slides) | David Blackwell | Han Zhou |
| 23/09/24 | Mechanistic Interpretability I (Slides) | David Blackwell | Ryan Chan |
| 02/10/24 | Mechanistic Interpretability II (Slides) | Delilah | Ryan Chan |
| 07/10/24 | Invited Talk: Federating Large Language Models from Scratch (Slides) | David Blackwell | Lorenzo Sani |
| 14/10/24 | Invited Talk: Causal Estimation of Memorisation Profiles (Slides) | David Blackwell | Pietro Lesci |
| 28/10/24 | No Language Left Behind (NLLB) Technical Report Overview (Slides) | David Blackwell | Giulia Occhini, Ryan Chan |
| 04/11/24 | Invited Talk: Ethnographic Approaches to AI Evaluations | Ursula Franklin | Jonas Kgomo |
| 18/11/24 | Biological neural networks (Slides, Slides) | David Blackwell | Balázs Mészáros , Jess Yu |
| 20/11/24 | Mechanistic Interpretability III (Slides) | Delilah | Ryan Chan |
| 25/11/24 | Application of foundation models in time series tasks | David Blackwell | Gholamali Aminian |
| 02/12/24 | Can language models play the Wikipedia game? (Slides) | David Blackwell | Alex Hickey, Jo Knight |
| 02/12/24 | Diffusion models | Ada Lovelace | James Thornton |
| 03/12/24 | Mechanistic Interpretability | Enigma | Neel Nanda |
| 09/12/24 | Scaling laws of neural networks (Slides) | David Blackwell | Edmund Dable-Heath |
| 16/12/24 | Improving training with better learning rate and batch size: Linear scaling rule from random matrix theory (Slides) | David Blackwell | Chanju Park |
Main
- Don't Count, Predict! paper
- Word Embeddings (1)
- Word Embeddings (2)
- Word Embeddings (3)
- Brief History of NLP (part 1)
- Brief History of NLP (part 2)
Extra
- Deep Learning, NLP and Representations
- Stanford NLP with Deep Learning | Lecture 1: Intro & Word Vectors
- Speech and Language Processing - Chapter 6: Vector Semantics and Embeddings
- Stanford Large Language Models | Lecture 1: Introduction
Main
- Neural Networks and Deep Learning | Chapter 1: Using neural nets to recognize handwritten digits
- Neural Networks and Deep Learning | Chapter 2: How the backpropagation algorithm works
- Alternatively, come along to Phil and Jack's Lunchtime Tech Talk on Back Propagation (April 11th) - message on the
#hut23-robots-in-disguiseslack if you want to get a calendar invite for the session
- Alternatively, come along to Phil and Jack's Lunchtime Tech Talk on Back Propagation (April 11th) - message on the
Extra
- Learning Deep Learning | Chapters 1 and 2
- Neural Networks by Hand | Feedforward Neural Networks
- Andrej Karparthy (YouTube): The spelled-out intro to neural networks and backpropagation
Main
- Speech and Language Processing | Chapter 9: RNNs and LSTMs
- Read up to Section 9.5 (read pages 1-13)
Extra
- The Unreasonable Effectiveness of Recurrent Neural Networks
- NLP with Deep Learning Stanford Course | Lecture 5
- Read up to Section 3 (read pages 1-11)
- Stanford NLP with Deep Learning | Lecture 5: RNNs)
Main
- Speech and Language Processing | Chapter 9: RNNs and LSTMs
- Read from 9.5 to 9.8 (read pages 14-21)
Extra
- Understanding LSTM Networks
- NLP with Deep Learning Stanford Course | Lecture 5
- Read from Section 3 (read pages 11 onwards)
- Stanford NLP with Deep Learning | Lecture 6: Simple LSTM RNNs
Main
Extra
- The Illustrated Transformer
- Andrej Karpathy's GPT-2 from scratch
- Anthropic's Transformer Circuits
- NLP with Deep Learning Stanford Course | Lecture 6
- Attention and Augmented Recurrent Neural Networks
- Stanford NLP with Deep Learning | Lecture 7: Translation, Seq2Seq, Attention
- Stanford NLP with Deep Learning | Lecture 9: Self- Attention and Transformers
- Michigan Deep Learning for Comp Vis | Lecture 13: Attention
Main
Extra
- Speech and Language Processing | Chapter 9: RNNs and LSTMs
- Read from 9.8 (read pages from 21 onwards)
- Speech and Language Processing | Chapter 10: Transformers and Pretrained Language Models
- NLP with Deep Learning Stanford Course | Self-Attention & Transformers
Main
Extra
- Speech and Language Processing | Chapter 11: Fine-Tuning and Masked Language Models
- The Illustrated BERT
- BERT 101 🤗 State Of The Art NLP Model Explained
- Paper summary — BERT
Main
Extra
- Paper summary - Improving Language Understanding by Generative Pre-Training
- Language Models are Unsupervised Multitask Learners
Note the below materials for other sessions, or are not confirmed
Main
- Gradient Based Learning Applied to Document Recognition
- An Image Is Worth 16x16 Words: Transformers for Image Recongition at scale
Extra
- Vision Transformer for Image Classification - Shusen Wang (YouTube)
- Neocognitron: A Self-organizing Neural Network Model for a Mechanism of Pattern Recognition Unaffected by Shift in Position
- But what is a convolution? - 3Blue1Brown (YouTube)
- CNN Explainer
Main
Extra
Main
Extra
Main
- RLHF: Reinforcement Learning from Human Feedback
- Illustrating Reinforcement Learning from Human Feedback (RLHF)
Extra
- An introduction to Reinforcement Learning
- Understanding Reinforcement Learning from Human Feedback (RLHF): Part 1
- Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback
- Secrets of RLHF in Large Language Models Part I: PPO
Beyond RLHF
- Reinforced Self-Training (ReST) for Language Modeling
- Direct Preference Optimization: Your Language Model is Secretly a Reward Model
Guides
Videos
Meta
Main
Extra
- Retrieval Augmented Generation: Streamlining the creation of intelligent natural language processing models
- Dense Passage Retrieval for Open-Domain Question Answering
- Lost in the Middle: How Language Models Use Long Contexts
- Self-RAG: Learning to Retrieve, Generate And Critique Through Self-Reflection
Main
Extras
There are plenty of blog-posts and top level overviews of diffusion models which explain the main idea of, 'running a noisy blurring process backwards from the noise', however for more technical reading (which I will warn are quite heavy on the maths) the main two papers are:
Both are about the sampling methods used in the process (notably without the inclusion of context that allows for text-to-image generation). For a general overview the following is fairly good:
And if you're curious (and want spoilers) about stable diffusion and latent diffusion models this is the main paper.
Main
Extra
- Large Language Models for Software Engineering: Survey and Open Problems
- A Systematic Evaluation of Large Language Models of Code
- Automated Program Repair in the Era of Large Pre-trained Language Models
- A Survey on Language Models for Code
Main
Main
Main
Extra
- Repo containing write-up and LaTeX source for the slides (may not be available immediately)
- lightning-GPT
- Brendan Bycroft's LLM Viz
- Hacked visualisaion for illustrating GPT-2 models
Main
Main
Main
Extra
Main
Extra
- Podcast: The 80,000 Hours Podcast on Artificial Intelligence
- Book: Nick Bostrom - "Superintelligence"
- Book: Toby Ord - "The Precipice"
Main
- (2020) Zoom In: An Introduction to Circuits
- (2021) A Mathematical Framework for Transformer Circuits
Supplementary
- CVPR tutorial: Intro to Circuits in CNNs by Chris Olah
- Transformer Circuits Playlist
- Neel Nanda's Walkthrough: A Mathematical Framework for Transformer Circuits
Extras
Main
- Combining Hierachical VAEs with LLMs for Clinically Meaningful Timeline Summarisation in Social Media
- Sig-Networks Toolkit: Signature Networks for Longitudinal Language Modelling
Extra
- Identifying Moments of Change from Longitudinal User Text
- ROFORMER: Enhanced Transformer with Rotary Position Embedding
- Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?
Expanding participatory governance for LLMs: case studies from BigCode, Aya Initiative, and Collective Intelligence Project
Main
- Towards Openness Beyond Open Access: User Journeys through 3 Open AI Collaboratives
- The BigCode Project Governance Card
- Aya Initiative: website, data paper
- Collective Intelligence Project white paper
Main
- 2022 in review: neuroAI comes of age
- Towards a Foundation Model of the Mouse Visual Cortex
- V1T: large-scale mouse V1 response prediction using a Vision Transformer
Main
- CodeMind: A Framework to Challenge Large Language Models for Code Reasoning
- Functional Benchmarks for Robust Evaluation of Reasoning Performance, and the Reasoning Gap
- The Counterfeit Conundrum: Can Code Language Models Grasp the Nuances of Their Incorrect Generations?
- THE GENERATIVE AI PARADOX: “What It Can Create, It May Not Understand”
Main
- The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits
- AI Transparancy Technique
- Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text
- Binoculars GitHub repo
Learn how to learn and distil during learning - Using meta-learning and second order optimisation to prune the model
Main
- Fisher-Legendre (FishLeg) optimization of deep neural networks
- Second order derivatives for network pruning: Optimal Brain Surgeon
- The Optimal BERT Surgeon: Scalable and Accurate Second-Order Pruning for Large Language Models
Main
Main
- HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal
- FINE-TUNING ALIGNED LANGUAGE MODELS COMPROMISES SAFETY, EVEN WHEN USERS DO NOT INTEND TO!
Main
Abstract It has now been a decade since the first adversarial examples were demonstrated on deep learning models. And yet, even still, we can not robustly classify MNIST images better than LeNet-5 or ImageNet images better than AlexNet. But now, more than ever, we need robust machine learning models. And not only robust to evasion attack: but also robust to poisoning, stealing, and many other attacks. In this talk I survey the current progress we have made on adversarial machine learning. While we have made many significant advances in making attacks practical, we have had made considerably less progress on defences. Making progress towards addressing these challenges will be of the highest importance in the coming years.
- ICASSP Tutorial
- Transformers primer
- Optimisation
- Approximation
- Memorisation
- In-context learning
We are in the midst of an ‘epidemic of online abuse’, which disproportionately affects women and minoritised groups. In recent years, technology companies and computer science researchers have made efforts to automate the identification of hate speech and other toxic or abusive language. However, existing resources are limited in a number of important ways, such as their lack of theoretical grounding and stakeholder input.The EPSRC funded project Equally Safe Online aims to harness stakeholder expertise to co-design resources and methods to tackle online GBV. In this talk, I will discuss outcomes and ongoing work from the project, focusing on participatory design for NLP, perspectivist approaches to dataset creation, and generation of counterspeech against hateful language.
When I started working on machine learning inference four years ago a typical model would run on a handful of CPU cores. We needed to think about distributing work between threads, but the systems-level problems and abstractions were well understood. Fast forward to today and machine learning models are so large that even a "small" language model can have billions of parameters and run across a multi-GPU system. In this talk I am going to go on an end-to-end journey through the implementation of these models. We will see some of the different problems which emerge in parallelism and distributed computing, and some of the places where I think we are lacking good abstractions.
- The Llama 3 Herd of Models
- GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints
- RoFormer: Enhanced Transformer with Rotary Position Embedding
- Mixture of Experts Explained - Huggingface blog
- Applying Mixture of Experts in LLM Architectures - Nvidia blog
Generative AI systems create risks which must be evaluated in order to be managed or mitigated. Current approaches to AI safety evaluation are primarily focused on assessing technical artefacts in isolation, and so may miss hazards that arise through human-AI-interaction or wide-scale deployment. In this talk, I introduce a sociotechnical approach to AI safety evaluation that aims to capture relevant complexity, to provide a more comprehensive safety assessment. In addition, different evaluation goals require matching evaluation methods. Reviewing the current landscape of AI safety evaluation, I point out strengths and key gaps that need to be addressed. I close by discussing trade-offs and open challenges at the frontier of AI safety evaluation research.
- Zoom In: An Introduction to Circuits
- A Mathematical Framework for Transformer Circuits
- Mapping the Mind of a Large Language Model
- Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
- Refusal in Language Models Is Mediated by a Single Direction
- Toy Models of Superposition
- Towards Monosemanticity: Decomposing Language Models With Dictionary Learning
- Scaling and evaluating sparse autoencoders
- Sparse Autoencoders Find Highly Interpretable Features in Language Models
Large language models (LLMs) offer unprecedented ML capabilities and continue to improve rapidly. As a result, various organizations are locked in a race to scale LLMs and explore their limits and weaknesses. We believe federated learning (FL) offers an untapped potential to dramatically increase the supply of data sources for these models. Early work has shown, for example, how LLM pre-training can tap into edge device data leveraging FL. Others have shown the impact of using federated optimizers in a poorly connected distributed infrastructure of stateful workers to train a centralized LLM.
We believe FL can reshape LLM practices and opportunities thanks to two of its most exciting features: relaxed synchronization requirements and privacy-by-design on users' data. The federated paradigm opens the doors of new interesting possibilities for the LLM community, like resource sharing, unbounded scaling on private data, democratization, and privacy. This talk contributes to the emerging field that blends the two worlds of FL and LLMs by presenting a fully federated approach for LLM pre-training from scratch. Our approach has shown to be viable at a scale of 3B parameters under a real working system.
In training language models, training choices—such as the random seed for data ordering or the token vocabulary size—significantly influence model behaviour. Answering counterfactual questions like "How would the model perform if this instance were excluded from training?" is computationally expensive, as it requires re-training the model. Once these training configurations are set, they become fixed, creating a "natural experiment" where modifying the experimental conditions incurs high computational costs. Using econometric techniques to estimate causal effects from observational studies enables us to analyse the impact of these choices without requiring full experimental control or repeated model training. In this talk, I will present our paper, Causal Estimation of Memorisation Profiles (Best Paper Award at ACL 2024), which introduces a novel method based on the difference-in-differences technique from econometrics to estimate memorisation without requiring model re-training. I will also discuss preliminary results from ongoing work that applies the regression discontinuity design to estimate the causal effect of selecting a specific vocabulary size.
- No Language Left Behind: Scaling Human-Centered Machine Translation (Report)
- Scaling neural machine translation to 200 languages (Nature paper)
- No Language Left Behind (NLLB) project
Two talks:
- Event-Based Learning of Synaptic Delays in Spiking Neural Networks
- Information-theoretic Analysis of Brain Dynamics & Neural Network Models Informed by Information Theory
- Toy Models of Superposition
- Towards Monosemanticity: Decomposing Language Models With Dictionary Learning
- Scaling and evaluating sparse autoencoders
- Sparse Autoencoders Find Highly Interpretable Features in Language Models
This project examines how Language Models can navigate Wikipedia. Which tests their ability to link semantically similar topics in a practical way. We have run experiments with a large variety of sentence embedding and large language models for comparison. We have also seen how the performance varies when transversing Wikipedia in other languages and when navigating between scientific papers, which allows an assessment of the breadth of the model's abilities.
Main