MIARA dossier thématique

Terence Broad

Item

Title

Terence Broad

Prénom

Terence

Nom

Broad

URL cv en ligne

https://terencebroad.com/

A comme Université

University of the Arts London

Photo de profil

terenceBroad_profilPhoto

Media

terenceBroad_profilPhoto

Linked resources

Filter by resource type and property: Items: All Items: Creator Items: director Items: Est cité par Items: Conférencier

Items with "Creator: Terence Broad"

Title	Class
Being Foiled	Image
Ghosts	Image
Teratome	Image

Items with "director: Terence Broad"

Title	Class
Blade Runner – Autoencoded	Film

Items with "Est cité par: Terence Broad"

Title	Class
« StyleGAN—you know the deepfakes of people—there's a website [called] ThisPersonDoesNotExist that's all using a GAN, using StyleGAN, and the way the GAN works is that we have these two networks, a generator and a discriminator. When I'm teaching about this I sort of introduce this like one person is a forger—that's the generator—and then the discriminator is a detective. So if someone's forging money, trying to make fake bank notes, then the discriminator is the person at the bank trying to figure out is this real money or is this fake money, and these are extremely effective. So the discriminator is trying to tell what's a real image compared to what's being generated, what's fake, and the generator's got this adversarial opposite objective which is: I want to generate things that trick the discriminator, passed off as being real. In this setup, the generator has one objective, and that objective is: I just need to trick this other network, I just need to get my bank notes passed off as real money, so to speak, or get my images of people or whatever passed off as being real, training data. »	author self citation
« Ultimately, I kind of came to this realization that what I was trying to do was I was trying to train a generative model without modelling data which is an oxymoron, it's a complete contradiction. It took me about a year to really figure this out but once I did, that was quite helpful, quite constructive, it made me rethink the problem, reframe the research I was trying to do, and so then I started thinking: "ok, we're not trying to train something from scratch to do something that isn't modelling data, how do we just push or nudge a general model away from an existing data set?" So how do we take something that's already been trained, and how do we kind of move it in a direction. And I ended up—instead of looking at very theoretical statistics in cognitive science—just looking at the actual structure of the models and really just thinking about their component parts. How can we play around with this? The talk is kind of an hacker's guide, and I started taking a hacker's mentality to working with generative AI where I was really just trying to pull these things apart, figure out what you can do, figure out the unintended ways that you could use them, figure out ways of using them that you weren't really supposed to do. »	author self citation

Items with "Conférencier: Terence Broad"

Title	Class
Artful Intrusions—A Hacker's Guide to Generative AI	Conference
Table ronde - Terence Broad, Kazushi Mukaiya, Gaëtan Robillard	Conference