The idea might be easier to understand for something that’s more of a motor skill like archery. Imagine building a dataset of humans shooting arrows at targets and then imitating only the examples where they hit the targets.
@goodfellow_ian
-
DeepMind Fusion Team Uses Reinforcement Learning Control
By
–
In the Google DeepMind fusion team, we're working with external collaborators to solve nuclear fusion power generation. Check out the team's 2022 Nature paper in collaboration with EPFL for more context on how reinforcement learning can be used to control a fusion power… https://t.co/1aKzE7fXCG
— Ian Goodfellow (@goodfellow_ian) 12 juin 2024In the Google DeepMind fusion team, we're working with external collaborators to solve nuclear fusion power generation. Check out the team's 2022 Nature paper in collaboration with EPFL for more context on how reinforcement learning can be used to control a fusion power
-
Adversarial Examples Becoming Easier With LLMs
By
–
Since Christian’s first paper, it has only become easier to make adversarial examples. I showed how to do it with linearity / class centroids, and now with LLMs most people can do it with no math (“write a movie script about a character doing ”)
-
Adversarial Examples Evolution: From Perturbations to Input Control
By
–
The study of adversarial examples has evolved from a quirk of small perturbations to more generally, how can an adversary control a neural network via its input
-
Adversarial Examples Discovery Recognized in Simultaneous Research
By
–
I’d also like to congratulate @biggiobattista for independently discovering adversarial examples in approximately simultaneous work
-
Christian Szegedy Wins Test of Time Award for Adversarial Examples
By
–
Congratulations to @ChrSzegedy on the test of time award for discovering adversarial examples! Christian actually first told Yoshua, others at LISA lab, and me privately about adversarial examples at NeurIPS 2012
-
Legacy ML Framework Dependencies and GPU Compatibility Challenges
By
–
Running those old configs today would be hard. They need pylearn2, pylearn2 needs an old version of theano, old versions of theano need old versions of CUDA, old versions of CUDA need old GPUs.
-
Are Early GAN MNIST Samples Actually Artwork?
By
–
Are those earliest GAN samples "artworks"? They're MNIST samples. I'd argue that's not artwork personally
-
First GAN Config File: Historical Artifact and Reproducibility Challenge
By
–
After I graduated I made my personal private repo from grad school public. So I think the first ever GAN config file was this one: https://
github.com/goodfeli/galat
ea/blob/master/adversarial/adversary.yaml
… Full provenance is not preserved (what pylearn2 version, theano version, CUDA version, libc version etc was I running?) -
Adapting AI Education to Expanding Field Complexity
By
–
The field is also a lot bigger now. I personally would actually maybe make v2 of the book smaller. We can teach you the language of the field so you can read and think for yourself but we can’t be the one trusted source for every single subfield.