Hattie Zhou: Lottery Tickets and Algorithmic Reasoning in LLMs

Hattie Zhou: Lottery Tickets and Algorithmic Reasoning in LLMs

Author: Daniel Bashir February 16, 2023 Duration: 1:42:59

In episode 60 of The Gradient Podcast, Daniel Bashir speaks to Hattie Zhou.

Hattie is a PhD student at the Université de Montréal and Mila. Her research focuses on understanding how and why neural networks work, based on the belief that the performance of modern neural networks exceeds our understanding and that building more capable and trustworthy models requires bridging this gap. Prior to Mila, she spent time as a data scientist at Uber and did research with Uber AI Labs.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

Subscribe to The Gradient Podcast:  Apple Podcasts  | Spotify | Pocket Casts | RSSFollow The Gradient on Twitter

Outline:

* (00:00) Intro

* (01:55) Hattie’s Origin Story, Uber AI Labs, empirical theory and other sorts of research

* (10:00) Intro to the Lottery Ticket Hypothesis & Deconstructing Lottery Tickets

* (14:30) Lottery tickets as lucky initialization

* (17:00) Types of masking and the “masking is training” claim

* (24:00) Type-0 masks and weight evolution over long training trajectories

* (27:00) Can you identify good masks or training trajectories a priori?

* (29:00) The role of signs in neural net initialization

* (35:27) The Supermask

* (41:00) Masks to probe pretrained models and model steerability

* (47:40) Fortuitous Forgetting in Connectionist Networks

* (54:00) Relationships to other work (double descent, grokking, etc.)

* (1:01:00) The iterative training process in fortuitous forgetting, scale and value of exploring alternatives

* (1:03:35) In-Context Learning and Teaching Algorithmic Reasoning

* (1:09:00) Learning + algorithmic reasoning, prompting strategy

* (1:13:50) What’s happening with in-context learning?

* (1:14:00) Induction heads

* (1:17:00) ICL and gradient descent

* (1:22:00) Algorithmic prompting vs discovery

* (1:24:45) Future directions for algorithmic prompting

* (1:26:30) Interesting work from NeurIPS 2022

* (1:28:20) Hattie’s perspective on scientific questions people pay attention to, underrated problems

* (1:34:30) Hattie’s perspective on ML publishing culture

* (1:42:12) Outro

Links:

* Hattie’s homepage and Twitter

* Papers

* Deconstructing Lottery Tickets: Zeros, signs, and the Supermask

* Fortuitous Forgetting in Connectionist Networks

* Teaching Algorithmic Reasoning via In-context Learning



Get full access to The Gradient at thegradientpub.substack.com/subscribe

Hosted by Daniel Bashir, The Gradient: Perspectives on AI moves beyond surface-level headlines to explore the intricate machinery and human ideas shaping artificial intelligence. Each episode is built on a foundation of deep research, leading to conversations that are both technically substantive and broadly accessible. You'll hear from researchers, engineers, and philosophers who are actively building and critiquing our technological future, discussing not just how AI systems work, but the larger implications of their integration into society. This isn't about speculative hype; it's a grounded examination of real progress, persistent challenges, and ethical considerations from those on the front lines. The discussions peel back layers on topics like model architecture, policy, and the fundamental science behind the algorithms becoming part of our daily lives. For anyone curious about the substance behind the buzz-whether you have a technical background or are simply keen to understand a defining technology of our age-this podcast offers a crucial and thoughtful resource. Tune in for a consistently detailed and nuanced take that treats artificial intelligence with the complexity it deserves.
Author: Language: English Episodes: 100

The Gradient: Perspectives on AI
Podcast Episodes
C. Thi Nguyen: Values, Legibility, and Gamification [not-audio_url] [/not-audio_url]

Duration: 1:30:13
Episode 127I spoke with Christopher Thi Nguyen about:* How we lose control of our values* The tradeoffs of legibility, aggregation, and simplification* Gamification and its risksEnjoy—and let me know what you think!C. Th…
Vivek Natarajan: Towards Biomedical AI [not-audio_url] [/not-audio_url]

Duration: 1:55:03
Episode 126I spoke with Vivek Natarajan about:* Improving access to medical knowledge with AI* How an LLM for medicine should behave* Aspects of training Med-PaLM and AMIE* How to facilitate appropriate amounts of trust…
Thomas Mullaney: A Global History of the Information Age [not-audio_url] [/not-audio_url]

Duration: 1:43:45
Episode 125False universalism freaks me out. It doesn’t freak me out as a first principle because of epistemic violence; it freaks me out because it works. I spoke with Professor Thomas Mullaney about:* Telling stories a…
Seth Lazar: Normative Philosophy of Computing [not-audio_url] [/not-audio_url]

Duration: 1:50:17
Episode 124You may think you’re doing a priori reasoning, but actually you’re just over-generalizing from your current experience of technology.I spoke with Professor Seth Lazar about:* Why managing near-term and long-te…
Suhail Doshi: The Future of Computer Vision [not-audio_url] [/not-audio_url]

Duration: 1:08:07
Episode 123I spoke with Suhail Doshi about:* Why benchmarks aren’t prepared for tomorrow’s AI models* How he thinks about artists in a world with advanced AI tools* Building a unified computer vision model that can gener…
Azeem Azhar: The Exponential View [not-audio_url] [/not-audio_url]

Duration: 1:46:25
Episode 122I spoke with Azeem Azhar about:* The speed of progress in AI* Historical context for some of the terminology we use and how we think about technology* What we might want our future to look likeAzeem is an entr…
David Thorstad: Bounded Rationality and the Case Against Longtermism [not-audio_url] [/not-audio_url]

Duration: 2:19:02
Episode 122I spoke with Professor David Thorstad about:* The practical difficulties of doing interdisciplinary work* Why theories of human rationality should account for boundedness, heuristics, and other cognitive limit…
Michael Sipser: Problems in the Theory of Computation [not-audio_url] [/not-audio_url]

Duration: 1:28:21
In episode 119 of The Gradient Podcast, Daniel Bashir speaks to Professor Michael Sipser.Professor Sipser is the Donner Professor of Mathematics and member of the Computer Science and Artificial Intelligence Laboratory a…