The Gradient: Perspectives on AI

chevron_right

Hattie Zhou: Lottery Tickets and Algorithmic Reasoning in LLMs

whatshot 43 snips

Feb 16, 2023

Guest

Hattie Zhou

01:42:59

forum

Ask episode

web_stories

AI Snips

view_agenda

Chapters

auto_awesome

Transcript

info_circle

Episode notes

insights

INSIGHT

Fortuitous Forgetting Enhances Learning

Fortuitous forgetting involves iterative forgetting undesirable info, then relearning desirable features.
This process improves model generalization by amplifying good information over cycles.

insights

INSIGHT

Connecting Grokking and Forgetting

Grokking may be explained as delayed generalization through iterative forgetting and relearning.
Training beyond interpolation is like simplifying the model by removing non-critical information, aiding generalization.

volunteer_activism

ADVICE

Embrace Iterative Refinement

Iterative refinement of models, like improving labels or retraining self-generated data, is a promising paradigm.
Embrace iterative training as a practical way to incrementally improve large models.

Get the Snipd Podcast app to discover more snips from this episode

What Got You Interested in Machine Learning?

01:45 • 3min

chevron_right

Is It a Good Idea to Be a Scientist?

04:56 • 3min

chevron_right

Empirical Theory, Understanding Deep Learning

07:34 • 2min

chevron_right

The Lottery Ticket Hypothesis

09:48 • 3min

chevron_right

Pre-Training and Initialization of a Lottery Ticket

13:16 • 3min

chevron_right

Rewinding a Lottery Ticket

16:29 • 6min

chevron_right

How Do You Set Weights to Zero?

22:04 • 4min

chevron_right

Is There a Way to Do Random Pruning?

26:27 • 2min

chevron_right

The Sign of Weights in the Super Mask

28:43 • 2min

chevron_right

Is the Basin of Attraction Really Large?

30:49 • 2min

chevron_right

How Much Precision Do You Need in the Weights?

32:46 • 5min

chevron_right

Is There a Difference Between a Weight Change and a Mask Search?

38:15 • 3min

chevron_right

Is Masking a Way to Stimulate Pre-Trained Large Language Models?

40:48 • 4min

chevron_right

Using Masking in Multitask Learning

45:02 • 2min

chevron_right

Can Fortuitous Remembering Be Used in Artificial Neural Networks?

47:28 • 3min

chevron_right

Retraining the Later Layers to Improve Transfer Learning Performance?

50:39 • 3min

chevron_right

The Relationship Between Training Loss and Grokking the Information Bottleneck Paper

53:48 • 4min

chevron_right

The Inductive Bias Picture Is Really Interesting

58:06 • 2min

chevron_right

Can We Develop Algorithms That Capture the Same Benefit Without Iterative Process?

59:43 • 4min

chevron_right

Teaching Algorithmic Reasoning via in-Context Learning

01:03:19 • 5min

chevron_right

The in-Context Learning Picture

01:08:22 • 5min

chevron_right

Induction Head Hypothesis - What Is Happening in in-Context Learning?

01:13:32 • 3min

chevron_right

In-Context Learning and Gradient Decading

01:16:25 • 5min

chevron_right

Is There a Strategy for Discovering Novel Algorithms?

01:21:31 • 3min

chevron_right

What's Next for You in Line With What You Began Investigating in This Paper?

01:24:41 • 2min

chevron_right

What I Saw at the Math and AI Workshop at NeurIPS

01:26:14 • 2min

chevron_right

Is There Still a Need for Scaling Regimens for Iterative Training?

01:27:45 • 5min

chevron_right

Is the Human Model Interaction Really Important?

01:32:50 • 5min

chevron_right

The Current Publishing Culture in ML

01:38:05 • 5min

chevron_right

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

Home Top podcasts Popular guests Top books