Ian Goodfellow, a well-known name in the field of artificial intelligence, who, among other things, came up with the concept of GAN, left Apple to push this discipline to the limit in DeepMind.
Every discipline has its celebrities, and their comings and goings can cause havoc throughout the industry. And Apple will probably make a bitter experience out of it; According to Bloomberg, he recently spilled the beans about a true AI wizard revolutionizing the field and slipped into competition with his team of top researchers under his arm. And the context of this change is surprising to say the least.
When a professional with a rare talent decides it’s time to change the stable, often this event keeps a large audience in suspense. Sports fans know this all too well, as they often face this situation during transfer windows.
One need only think of the arrival of Lionel Messi at PSG, the frenzy that surrounded Kylian Mbappe’s decision, or the huge media storm caused by the departure of LeBron James when he moved his talents from Cleveland to Miami to see this; the thunderclap passed to the descendants called “Decision”. And recently, the AI world has potentially just experienced a comparable episode with Ian Goodfellow, an eminent researcher who is about as famous in his field as the superstars above.
Academic superstar with a rich resume
The stakeholder is a real academic war machine. Together with his team, they have already prepared a number of research papers that redefine some aspects of AI down to their foundations; one need only look at the academic search engine Google Scholar to understand this. For reference, in the scientific field, the impact and coverage of certain papers can be (roughly) estimated by looking at the citation count, that is, the number of times other research has been based on that work.
In most fields, a few hundred citations in other papers are often an important guarantee of quality; this shows that the article is not only serious, but also cutting-edge research in its field. In the case of Goodfellow, only his three most cited publications (here, here and here) have… more than 100,000 citations.
An absolutely insane number that says a lot about his defining influence in the AI world; every day, researchers in the four corners of the planet rely directly on his work to advance the discipline even further.
The Father of Generative Adversarial Networks
If Goodfellow has such influence, it is partly because he was the first to imagine and then flesh out the concept of Generative Adversarial Networks (GANs) in 2014. In short, this is a somewhat peculiar version of the standard neural network – those tangles of logical units that support what is commonly called “artificial intelligence”.
The concept was simply revolutionary when it was introduced in 2014. The peculiarity of these GANs is that they consist of two subnets that compete with each other – hence the term “adversarial”. The first, called the “generator”, creates a sample with the aim of making his opponent – the “discriminator” – believe that this is a real image. Each time, the discriminator becomes harder to fool. Therefore, the former should give more and more convincing results until it reaches a stage where people are completely unable to tell the difference.
Diagram of the GAN concept. © Google Developers
This is the approach that caused a small revolution in AI; now there is no need to customize each iteration to the smallest detail. In GAN, by definition, the discriminator is responsible for sorting upstream, which makes life much easier for researchers in certain fields and allows you to produce incredible results when it comes to generating an image or a piece of music.
For this reason, today they are used in many art-related applications. All those photorealistic AI-generated faces you’ve probably seen on the internet are, for example, the result of GANs like GauGAN, Nvidia’s amazing artistic AI. (see our article). And this technology is not limited to art and deepfakes. Its main points can also be used in areas such as neuroscience, pharmaceutical research, etc.
Getting fired due to… remote work
Suffice it to say that the genius behind this concept was quickly courted by the titans of the industry. He made his debut with great fanfare, successively joining the Google Brain Labs, OpenAI, then Google Research. Then he changed course, becoming the director of Apple’s powerful machine learning department, with the results we know.
Obviously, it would be unfair to credit him with all of Apple’s success in this segment; Apple already had great engineers and a solid logistics base. But Goodfellow’s arrival was still a heavy burden. It is no coincidence that the firm is today one of the world leaders in hardware optimization thanks to AI.
But, in spite of everything, this idyll ended in April last year for reasons, to put it mildly, curious. Indeed, if an interested party has packed, it is not a matter of salary or project; this is simply because he flatly refused to follow the return-to-work policy introduced by Apple in the context of slowing down the Covid-19 pandemic.
Disagreements over remote work were the reason for this departure. © Andrea Piacquadio – Pexels
Were there other underlying reasons? Very likely, because sacrificing such a size on the altar of such a policy seems simply unthinkable. Especially in a field like artificial intelligence, where working from home is quite manageable.
If you draw a parallel with the world of sports, it’s a bit like a Champions League finalist voluntarily missing out on a potential Ballon d’Or in the middle of the Champions League final phase on the pretext that he would have landed badly on the pitch. luggage storage; this is a surprising explanation that suggests something fishy is going on.
Focus on general purpose AI?
But whatever the root of the problem, the result is the same: Goodfellow is no longer an Apple employee, and he was immediately approached from all sides. He ended up at DeepMind, another member of the global AI elite run by Alphabet, Google’s parent company, according to Bloomberg. To seduce him, she most likely gave him carte blanche for a remote job; but there is no doubt that she put forward other very tempting arguments.
We don’t yet know exactly what Goodfellow will be working on at DeepMind. Recall that the company is engaged in very diverse areas of application, where it often offers quite revolutionary work. We can refer to AlphaFold, a database released in July 2021 that has already redefined structural biology (see our article). On the other hand, it is unlikely that Goodfellow will work on these rather specialized projects. He will likely pursue the long-term goal of DeepMind.
© Geralt – Pixabay
As a reminder, the firm aims to develop a so-called “universal” AI, as opposed to specialized neural networks that exist today. Thus, such a system will be able to solve problems that are too complex for the human brain, and a real social revolution is at stake. DeepMind has also shown progress in this direction very recently with its Gato system, even if the term “generalist” AI is still taken with a grain of salt (see our article).
The promise of fundamental research work of exceptional scope with a philosophy quite different from Apple’s, which puts AI first and foremost at the service of hardware optimization and user experience. Such an argument, to which such advanced specialists as Goodfellow are probably not indifferent. Thus, it seems that the alliance between a great specialist in this discipline and an AI titan with an excessive appetite makes sense.
This is a stable change that could have important implications in a few years. These moves are certainly frequent in the world of big tech, but it’s not every day that a heavyweight of this caliber switches sides. This is certainly an excellent catch for an institution that is already part of the global elite, and therefore a sacred loss for Apple. The firm may well bite its fingers if a stakeholder ends up contributing to the development of the first general-purpose AIs… all for the sake of a vulgar story about a physical presence.