The Introduction of Transformers as Programmable Computers

4min Snip

00:00

Play full episode

Summary

Transcript

Episode notes

In this talk you gave at the Alan Turing Institute you kind of expounded on how RNNs relate to the models of computation we're familiar with. You made the argument that in the hierarchy they lie closer to finite state machines I don't necessarily have access to unbounded memory than to Turing machines. How do you think about the introduction of transformers in terms of that computational hierarchy question? That's an interesting question to ask specifically because I had one of my student Minchi Jiang sent me a paper just the other day on this topic is called Looped Transformers as Programmable Computers by Angelique Genu and colleagues.

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