Transformer Circuits, Induction Heads, In-Context Learning

Name: Transformer Circuits, Induction Heads, In-Context Learning
Uploaded: 2022-07-17T21:00:33.000Z
Duration: 59 min 34 s
Description: This lecture explores mechanistic interpretability in transformer language models, focusing on "induction head circuits." It elucidates how these attention head pairs enable in-context learning by allowing models to repeat or translate text, and mimic functions from prior examples.

Ashish Vaswani; Lucas Beyer; William Fedus

Browse

Prev Next

Stanford University·Ashish Vaswani, Lucas Beyer & William Fedus

59:34

0:00/59:34

Mechanistic Interpretability, Alignment & World Models →

0/6

01

The Deep Learning Revolution (Turing Award Lecture)

Geoffrey Hinton · 32:28

This foundational lecture provides the historical context and conceptual genesis of deep learning, setting the stage for understanding the complex systems we aim to interpret and align.

02

Transformer Circuits, Induction Heads, In-Context Learning

Ashish Vaswani, Lucas Beyer & William Fedus · 59:34

Building on the deep learning revolution, this lecture introduces mechanistic interpretability by dissecting transformer circuits, specifically induction heads, to understand how models achieve in-context learning.

03

Objective-Driven AI: World Models that Learn, Remember, Reason, and Plan

Yann LeCun · 1:16:53

Expanding beyond token prediction, this step introduces objective-driven AI and world models, offering a framework for how AI systems learn, remember, reason, and plan, which is crucial for alignment.

04

Atomic Facts to Structured Knowledge: Rethinking Unlearning & Jailbreaking in Large Language Models

Rongzhe Wei · 56:22

With an understanding of world models, this lecture delves into structured internal knowledge modeling, which is essential for addressing unlearning and jailbreaking, directly impacting AI safety and alignment.

05

AI and Safety

Ron Rivest · 29:26

This lecture directly addresses critical aspects of AI safety, exploring methods and theoretical frameworks to ensure robust and reliable operation, building upon the technical insights from interpretability and world models.

06

Governance at the Technological Frontier: Translating AI Research into Oversight Policy

Inioluwa Deborah Raji · 39:55

The final step translates technical AI safety research into practical oversight policy, examining the institutional and regulatory mechanisms needed to govern advanced AI systems effectively.

Transformer Circuits, Induction Heads, In-Context Learning

Ashish Vaswani Lucas Beyer William Fedus Stanford University

59:34·2022·graduate level

Path step 2: Building on the deep learning revolution, this lecture introduces mechanistic interpretability by dissecting transformer circuits, specifically induction heads, to understand how models achieve in-context learning.

This lecture explores mechanistic interpretability in transformer language models, focusing on "induction head circuits." It elucidates how these attention head pairs enable in-context learning by allowing models to repeat or translate text, and mimic functions from prior examples.

"Neural network parameters can be thought of as compiled computer programs. Somehow, they encode sophisticated algorithms, capable of things no human knows how to write a computer program to do. Mechanistic interpretability seeks to reverse engineer neural networks into human understandable algorithms. Previous work has tended to focus on vision models; this talk will explore how we might reverse engineer transformer language models.

AI & ML Language Models & NLP Mechanistic Interpretability, Alignment & World Models Mechanistic Interpretability & Neural Circuits Large Language Models (LLMs)Transformers & Attention Neural Networks & Backpropagation

Study path·Step 2 of 6

Mechanistic Interpretability, Alignment & World Models

View full sequence →

Previous · Step 1

The Deep Learning Revolution (Turing Award Lecture)

Next · Step 3

Objective-Driven AI: World Models that Learn, Remember, Reason, and Plan

Lecturers

3

Speaker directory →

Ashish Vaswani

Co-Creator of the Transformer Architecture ('Attention Is All You Need')

Machine learning researcher and co-first author of 'Attention Is All You Need', introducing the self-attention Transformer architecture that underpins modern foundation models.

1 course · 9 lectures in archiveView profile

Lucas Beyer

Co-Creator of Vision Transformer (ViT), Google Brain & OpenAI

Computer vision and deep learning researcher who co-created the Vision Transformer (ViT), SigLIP, and large-scale multimodal representation models.

1 course · 9 lectures in archiveView profile

William Fedus

Professor at Stanford University

William Fedus is a professor at Stanford University, specializing in machine learning and the development of transformer models and mixture of experts.

1 course · 9 lectures in archiveView profile