Slashdot Mirror
Apple Hires AI Expert Ian Goodfellow (cnbc.com)
One of Google's top minds in artificial intelligence has joined Apple in a director role. Ian Goodfellow said on his LinkedIn profile that he switched employers in March. He said he's a director of machine learning in the Special Projects Group. CNBC reports: Goodfellow is the father of an AI approach known as generative adversarial networks, or GANs. The approach draws on two networks, one known as a generative network and the other known as a discriminative network, and can be used to come up with unusual and creative outputs in the form of audio, video and text. GAN systems have been used to generate "deepfake" fake media content.
Goodfellow got his Ph.D. at the University of Montreal in 2014, and since then he has worked at OpenAI and Google. At OpenAI he was paid more than $800,000, according to a tax filing. His research is widely cited in academic literature. At Google Goodfellow did work around GANs and security, including an area known as adversarial attacks. People working on AI at Apple have previously done research that drew on the GAN technology.
Goodfellow got his Ph.D. at the University of Montreal in 2014, and since then he has worked at OpenAI and Google. At OpenAI he was paid more than $800,000, according to a tax filing. His research is widely cited in academic literature. At Google Goodfellow did work around GANs and security, including an area known as adversarial attacks. People working on AI at Apple have previously done research that drew on the GAN technology.
I know it's from TFA's headline, but GFA is "Generative Adversarial Networks", not what they said.
Also very good at marketing himself although he build his 'unique' expertise on the work of others.... https://en.wikipedia.org/wiki/... This is nothing more than a pawn in a mud fight between Google and Apple over the 'best' engineers.
Is it just me, or does "AI expert Ian Goodfellow" sound like a character from a William Gibson novel?
You are welcome on my lawn.
Not so much any more. This is solid indication of bailout of non-violent persons who have no interest in working for Death Incorporate, the US War Industrial Complex, are bailing on Google, the technologists of Death Incorporated, whilst trying to skull behind ethicsvertising and outed by Donald Trump himself. Google don't be evil, FOR FREE.
A lot more Googlites will be seeking to bail out the evil Google and recruiting from there will be a lot easier. Hardly surprising Apple is now able to pick the best from Google.
Chaos - everything, everywhere, everywhen
he build his 'unique' expertise on the work of others...
Hey you know who else did that?
If I have seen further than others, it is by standing upon the shoulders of giants.
-Issac Newton.
Hey you know who ELSE has done that?
Anyone who has ever used StackOverflow to find an answer.
Don't knock building expertise on what others before you have done.
"There is more worth loving than we have strength to love." - Brian Jay Stanley
Somehow 5 years later he is an "industry leader in machine learning"?
That seems very doable to me, given how recent a lot of the modern approaches to machine learning are... that is in fact about the right time that if you went into applying machine learning around five years ago, you could easily be a leading expert at this point.
My thought is, the proof is in the hire. Lots of people here spouting a lot of trash but two high level companies agree he's an expert. That sounds a lot more informative than AC trash-talk to me...
"There is more worth loving than we have strength to love." - Brian Jay Stanley
Somehow 5 years later he is an "industry leader in machine learning"?
Perhaps it had something to do with publishing original research that revolutionized machine learning.
Remember hearing about Alphazero defeating the world's Go champion? That was based on a GAN, which Ian invented.
He also wrote the definitive book on deep learning which anyone in the field should read. Twice.
What is special about GANs over and above making your AI's co-evolve through competing with each other, which is nothing new? (I'm not being snarky, I do presume there is more to it).
I think the reason AI is suddenlt leaping forward after being stallaed for decades is
1. almost entirely because computers are faster now. Especially because GPUs are faster. And specifically because of all machine learning algorithms out there the one that maps to matrix multiplies most easily is the Neural Net. It was just waiting for a graphics card with enough speed and memory to reach a practical scale.
2. Almost all mchachine learning is a refinement of past ideas. But GANs, seem to me to be the big breakthrough. They are uncanny.
In fact the thing GANS remind me most of is Dreaming in sleep. In dream time we replay ideas and try to conjur up alternatives. Part of our mind is just generating images and the other part is trying to classify them so it can reject the ones that can't fit a narrative. It really makes me think that GANs may have unlocked the secret of human intelligence of learning in this generative adversarial way that lets you generalize from fewer examples to rules. This could be hogwash. But GANs themselves are a huge breaktrhough
Some drink at the fountain of knowledge. Others just gargle.
In regular co-evolution you'd generally have a cost/fitness function to judge fitness by. Evolutionary algorithms are essentially optimizers that do well with non-convex problems.
GANs were invented to answer the question "what do you do when you don't know what the cost function is?" in the particular circumstance when you want to build a generative model. You want the output of your generative model to be as much as possible like a true example, but it's difficult to properly measure similarity. There are lots of similarity metrics, but they all tend to concentrate on particular features, and don't work terribly well for generation.
So in a GAN you let a second neural network *learn* what the cost function is, based on the simple objective: decide which of these examples are real and which were generated. You train both models together with the idea that the generative model has to figure out how to create examples that are enough like the real ones that the critic can't find a way to tell them apart.
As you might imagine, the whole thing depends on the capacity of the critic, and GANs have been quite difficult to train because the generator is very good at exploiting limitations in the the critic.
Thank you. Although my understanding is that co-evolution is defined as having no fixed evaluation function, but rather selection is based on viability within an environment with one or more other species that are also evolving.
That's a good point. I think selection by viability implies some kind of fitness function embedded in the environment, but it doesn't have to be fixed, simple, or explicit.
I suspect you could look at GANs in a few different ways in a co-evolutionary framework. Goodfellow presents them as adversarial (it's even in the name) and gives the example of an art forger (the generator) and an art authenticator (the critic). That's kind of a predator-prey setup. Personally, I think a better way to look at it is more of a student-teacher relationship, with the teacher having to do some learning on the job to keep up with a bright student.
One of the big differences between a GAN and a typical evolutionary-type experiment is that there's only one generator and one critic, both of which are updated via gradient descent. In an evolutionary setup normally you'd have multiple individuals competing and optimization would be done via selection.
Deep neural networks are special because of how flexible they are. Many of the really cool applications, like reinforcement learning (e.g. AlphaGo) have been around forever but when you stick a DNN in you suddenly get something really useful. I expect there's going to be a lot of interesting work involving DNNs and evolutionary algorithms. I think you would probably arrive pretty naturally at a GAN via combining DNNs and co-evolution.