I'd Like to Introduce a Few of the Main Contributions in This Paper

5min Snip

00:00

Play full episode

Summary

Transcript

Episode notes

I do think that the data side of training these models, I mean, people have already looked at scaling laws and how much data. But as you said, there's been not as much focus on constructing the data sets. So I definitely am excited to see follow and work to yours. Let's perhaps move to some of your other work. And we can maybe bridge the jumping around we've done time wise. You had to work on a convolutional neural network for modeling sentences, which I thought was a pretty interesting paper.You included this new development dynamic kmax pooling, where you observed a problem in this previous model max to DNN and resolved it with this idea. It

In episode 62 of The Gradient Podcast, Daniel Bashir speaks to Ed Grefenstette.

Ed is Head of Machine Learning at Cohere and an Honorary Professor at University College London. He previously held research scientist positions at Facebook AI Research and DeepMind, following a stint as co-founder and CTO of Dark Blue Labs. Before his time in industry, Ed worked at Oxford’s Department of Computer Science as a lecturer and Fulford Junior Research Fellow at Somerville College. Ed also received his MSc and DPhil from Oxford’s Computer Science Department.

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

* (02:18) The Ed Grefenstette Origin Story

* (08:15) Distributional semantics and Ed’s PhD research

* (14:30) Extending the distributional hypothesis, later Wittgenstein

* (18:00) Recovering parse trees in LMs, can LLMs understand communication and not just bare language?

* (23:15) LMs capture something about pragmatics, proxies for grounding and pragmatics

* (25:00) Human-in-the-loop training and RLHF—what is the essential differentiator?

* (28:15) A convolutional neural network for modeling sentences, relationship to attention

* (34:20) Difficulty of constructing supervised learning datasets, benchmark-driven development