" Accessible Foundation Models: Systems, Algorithms, and Science"

Abstract

The ever-increasing scale of foundation models, such as ChatGPT and AlphaFold, has revolutionized AI and science more generally. However, increasing scale also steadily raises computational barriers, blocking almost everyone from studying, adapting, or otherwise using these models for anything beyond static API queries. In this talk, I will present research that significantly lowers these barriers for a wide range of use cases, including inference algorithms that are used to make predictions after training, finetuning approaches that adapt a trained model to new data, and finally, full training of foundation models from scratch.  For inference, I will describe our LLM.int8() algorithm, which showed how to enable high-precision 8-bit matrix multiplication that is both fast and memory efficient. LLM.int8() is based on the discovery and characterization of sparse outlier sub-networks that only emerge at large model scales but are crucial for effective Int8 quantization. For finetuning, I will introduce the QLoRA algorithm, which pushes such quantization much further to unlock finetuning of very large models on a single GPU by only updating a small set of the parameters while keeping most of the network in a new information-theoretically optimal 4-bit representation. For full training, I will present SWARM parallelism, which allows collaborative training of foundation models across continents on standard internet infrastructure while still being 80% as effective as the prohibitively expensive supercomputers that are currently used. Finally, I will close by outlining my plans to make foundation models 100x more accessible, which will be needed to maintain truly open AI-based scientific innovation as models continue to scale.

 
Biography

Tim Dettmers’ research focuses on making foundation models, such as ChatGPT, accessible to researchers and practitioners by reducing their resource requirements. This involves developing novel compression and networking algorithms and building systems that allow for memory-efficient, fast, and cheap deep learning. These methods enable many more people to use, adapt, or train foundation models without affecting the quality of AI predictions or generations. He is a PhD candidate at the University of Washington and has won oral, spotlight, and best paper awards at conferences such as ICLR and NeurIPS. He created the bitsandbytes library for efficient deep learning, which is growing at 1.4 million installations per month and received Google Open Source and PyTorch Foundation awards.

"Paths to AI Accountability"

Abstract

We have begun grappling with difficult questions related to the rise of AI, including: What rights do individuals have in the age of AI? When should we regulate AI and when should we abstain? What degree of transparency is needed to monitor AI systems? These questions are all concerned with AI accountability: determining who owes responsibility and to whom in the age of AI. In this talk, I will discuss the two main components of AI accountability, then illustrate them through a case study on social media. Within the context of social media, I will focus on how social media platforms filter (or curate) the content that users see. I will review several methods for auditing social media, drawing from concepts and tools in hypothesis testing, causal inference, and LLMs.

 
 Biography

Sarah is a final-year PhD student at MIT in the Electrical Engineering and Computer Science Department advised by Professor Aleksander Mądry and Professor Devavrat Shah. Sarah utilizes methods from machine learning, statistical inference, causal inference, and game theory to study responsible computing and AI policy. Previously, she has written about social media, trustworthy algorithms, algorithmic fairness, and more. She is currently interested in AI auditing, AI supply chains, and IP Law x Gen AI.

Research Interests/Area

Responsible AI, ethics in AI, AI policy + statistical inference, causal inference

"Understanding Language Models through Discovery and by Design"

Abstract

Whereas we understand technologies like airplanes or microprocessors well enough to fix them when they break, our tools for fixing modern language models are coarse. This is because, despite language models' increasing ubiquity and utility, we understand little about how they work. In this talk, I will present two lines of research for developing a deep, actionable understanding of language models that allows us to discover how they work, and fix them when they fail. In the first line, I will present structural probing methods for discovering the learned structure of language models, finding evidence that models learn structure like linguistic syntax. In the second line, I will show how we can understand complex models by design: through the new Backpack neural architecture, which gives us precise tools for fixing models.

 
 Biography

John is a PhD student in Computer Science at Stanford University, working with Percy Liang and Christopher Manning on discovering the learned structure of neural language models, and designing them to be more understandable, diagnosable, and fixable. He was an NSF Graduate Fellow, and received a B.S.E in Computer and Information Science from the University of Pennsylvania. John has received an Outstanding Paper Award at ACL 2023, a Best Paper Runner Up at EMNLP 2019, an Honorable Mention for Best Paper at the Robustness of Few-Shot Learning in Foundation Models Workshop (R0-FoMo)@NeurIPS 2023, and an Outstanding Paper Award at the Workshop on Analyzing and Interpreting Neural Networks for NLP (BlackBoxNLP)@EMNLP 2020.

Research Interests/Area

Natural Language Processing

CS Department Events

CSPAC Workshop Series

Other Events & Opportunities

Northwestern Medicine Healthcare AI Forum

3rd Annual Traditional Spring Pow Wow- Hosted by NAISA

CS News

NU News