Forget Deep Learning and AlphaGo. The next phase in AI is growing behind the scenes, and it’ll put Darwin in our computers… going far beyond artificial neural networks.

It’s 9 AM and the sun warms the San Francisco bay. A man in his 50s, toying with his blond hair, sits down to his computer and connects to Zoom, a videoconferencing app. Someone from Paris asks why he left the University of Texas at Austin in 2015, where he’d taught computer science and neuroscience for 28 years, to join Sentient Technologies. The Finnish Risto Miikkulainen grins: “I’ve been leading research into artificial intelligence, and specifically into evolutionary algorithms, since 1990. But for years, I’ve wanted more than theory. I wanted to give them a concrete application, with the idea that evolutionary computation was ready to be commercialized and to change the world.

On the 23rd floor of One California, an impressive building in San Francisco’s business district overlooking the Golden Gate Bridge, Sentient was cofounded three years ago by the French Antoine Blondeau. Risto Miikkulainen’s team, consisting of three students and four researchers, is now working diligently to finalize a huge artificial intelligence platform that could fix highly complex problems requiring massive amounts of data, particularly in e-commerce, trading, and medical research.

Risto Miikkulainen is one of the pioneers of “neuroevolution,” a kind of AI that uses evolutionary algorithms (EA) to generate artificial neural networks (ANN), used especially in robotics. But what is an algorithm relying on evolutionary computation, exactly? It’s a relatively unknown branch of AI that uses the theory of evolution, and more specifically the principle of Darwinian natural selection, to optimize or fix complicated problems.

To put it simply, evolutionary algorithms imitate the capacity of a population of living organism to adapt themselves to their environments to survive, using selection mechanisms and genetic inheritance. They randomly generate a “population” of solutions (a set of parameters, similar to a cooking recipe) to a given problem. Then the best solutions are combined to design “descendants,” and the process carries itself out from generation to generation until the best possible solution is found.

Darwin in circuits

The idea is not to completely replicate biological evolution, only to transpose some of its major principles to computer science – the growth of genetic material, mutation, and the selection of variations – in order to fix problems much more easily and quickly,” Risto Miikkulainen explains. Thanks to evolutionary computation, Sentient’s researchers designed an algorithm three years ago that can, without any outside aid, fix any kind of problem and adapt to its environment. “We go first through deep learning (DL) by simulating neural networks, in order to observe and analyze huge amounts of data, and then create a modeling of the world. But to make decisions, we’re turning to evolutionary computation, because DL makes it difficult to make decisions,” the scientist says.

Ever since Lee Sedol, master of the game Go, lost two years ago to AlphaGo, Google’s DeepMind artificial intelligence, deep learning has captured the public’s attention on the subject of AI progress. This technology, based on artificial neural networks, aims to imitate human intelligence as much as possible. But it has its limits.

In the University of Strasbourg’s ICube lab, a white building that looks like a spaceship, the Complex Systems and Translational Bioinformatics department (CSTB) houses 30 researchers working on artificial “bio-inspired” systems. Among them, Pierre Collet, himself a specialist in evolutionary computation (he prefers the term artificial evolution,” or EA) has been working for 20 years on the design of “evolutionary motors.” More or less the same type of platform developed by Sentient, they would allow computers to “create industrial objects that perform better than those created by engineers” by simulating Darwinian evolution in a machine.

Professorial without relying on opaque jargon, the scientist explains that EA’s goal is simply to make computers creative – which is not the goal of deep learning: “By simulating millions of neurons, connected and arranged in hundreds of different layers, a machine is capable of analyzing data, learning and understanding its environment, for example by recognizing images. But it won’t be able to find new solutions to a problem, because it’s based on exploiting things we already know. That’s where EA comes into play, because by following a creative process, it allows a computer to find new solutions without human help.” The solutions found by artificial evolution machines are sometimes surprising and beyond our comprehension, such as these plans for a mini satellite antenna designed by computers in 2006 for NASA: “Human engineers could never have imagined an antenna like this.

According to Pierre Collet, “evolutionary algorithms are already regularly producing results, patented, as productive as human engineers, even better.” EA is used to resolve concrete problems, from control of gas pipeline flow to the design of airplane wing profiles, as well as air routing, creating schedules at universities, network routing design, and even planning robot trajectories. “Recently, I worked with a company that manages a fleet of medical vehicles, in order to use EA to figure out how to optimize use. Any industrial problem can be resolved by artificial evolution, because it’s a universal fixer,” the researcher says.

Second artificial revolution

San Francisco, California Street. While Cable Cars, the famous tramways in the “City by the Bay,” follow one another down the street, Sentient’s researchers have just made several breakthroughs. First in the solving of extremely difficult computation problems: “Our evolutionary computation platform allows us to resolve problems with up to a billion variables,” Risto Miikkulainen explains, downing his coffee. Then, in terms of real-world optimization of applications: “Our evolutionary motor is able to produce new designs for website designers that humans couldn’t even imagine.

Finally, with its evolutionary computation platform, Sentient could improve deep learning architecture (sprawling and more and more complex to design by human engineers). “Our goal is to allow manufacturers to optimize and improve their learning systems beyond human capacities: AI builds AI, and the architectures of DL evolve by themselves,” Risto Miikkulainen says. On the company’s site, several demos allow users to visualize the possible uses of such an artificial evolution: “Music Maker,” a system of modeling language, has seen its memory structure (Long Short-Term Memory, or LSTM) increase tenfold by EA, and Celeb Match, a program for recognizing images, has seen its own network considerably improved by evolutionary computation.

ICube’s and Sentient’s team are pursuing the same goal: to construct algorithms capable of attacking any kind of problem all the way to its end, without human intervention. “To build AlphaGo, Google hired between 10 and 20 scientists for three years. Our idea is to replace the human part with an intelligent and evolving system that works on its own on thousands of graphics cards and processors,” Antoine Blondeau, founder of Sentient, said last year to Numerama. “The kind of paring between DL and EA doesn’t exist yet, but without a doubt it’s the next step in AI,” Risto Miikkulainen says, shaking his head.

While Sentient seeks to use EA to optimize deep learning, Pierre Collet tries to use artificial evolution to fix problems initially identified by deep learning – the idea being to use DL to identify a difficult problem, then to launch an evolutionary algorithm to resolve it creatively. “Concrete applications would be infinite. Although DL lets you imitate compositions, EA could create entirely new musical pieces. By designing an autonomous computer able to understand its environment, to be inventive and to resolve problems on its own, you could also imagine a robot on Mars managing to move a rock in front of it by deciding to use its arms to manipulate it, even though they weren’t built to move obstacles,” Pierre Collet says. “We’ve arrived at a key moment in the history of AI. Evolutionary computation, combined with deep learning, will be able to generate totally new solutions,” Risto Miikkulainen adds.

Progressing little by little, Sentient’s researchers are concentrating at the moment on industrial and creative fields like marketing, web design, e-commerce, and cybersecurity: their evolutionary algorithms, for example, let you create an AI capable of connecting clothing photographed in the street to offers on online shops, or even to identify network security errors.

But other projects could also save lives. The AI company based in San Francisco, with the help of MIT, is designing a system that could predict seismic shocks. “In the field of agriculture, we could use evolutionary algorithms to optimize crop yields and to transform the future of food production. Finally, in medical research, one could use the genetic scissors CRISPR in a much more precise way,” Risto Miikkulainen says. In Strasbourg, Pierre Collet is also working with his evolutionary motors to let small drones “sneak between debris after an earthquake and to decide by itself what needs to be done when it finds survivors: like locating them and identifying obstacles that will let emergency responders act more effectively.

Another form of intelligence

“The challenge is huge,” Risto Miikkulainen says. At the moment, evolutionary algorithms, less handsome than deep learning when it comes to visual applications, remain relatively unknown to the media and the general public. And yet they’ve been studied since the beginnings of AI. The first patents for EA solutions and for “artificial Darwinism” came at the beginning of the 2000s with the rise of machine power.”This technology is so effective that more and more companies are choosing to use EA engineers capable of resolving all kinds of problems, instead of a multitude of engineers specialized in a single field,” Pierre Collet notes.

Besides, other major actors in AI also seem interested in artificial evolution, starting with those in Silicon Valley who, until now, swore only by deep learning. At Google, artificial intelligence researchers are exploring evolutionary algorithms for the classification and research of images, a complement to deep learning, or using the EA of DeepMind to better train artificial neural networks.

Scientists at the OpenAI project, cofounded by Elon Musk in 2015 to “develop a human-like AI that will benefit all humanity,” are trying to use evolutionary computation to “teach computers to perform particularly complex tasks.” According to its American researchers, the “strategies of evolution” (a branch of artificial evolution born in Germany in 1965) could even be “an alternative to learning by reinforcement,” a method of machine learning that consists of teaching an algorithm to use observations on actions in its environment… and which was the origins of AlphaGo.

Together, Pierre Collet says that deep learning and artificial evolution could finally lead to “hard” or “strong” artificial intelligence; in other words, a truly intelligent AI. In Strasbourg, among a bank of computers, the French researcher is excited about the idea. “That’s the future, for me, and that’s what I’m working for: to finally design autonomous computers that identify problems and design good solutions on their own. For now, AI is still weak, with computers that can recognize their environment but which don’t know what to do with it. EA promises to change that.

But the “intelligence” of this new type of computer will not really be human. “Autonomous computers might seem like fantasy, but it will only be so as long as we try to reproduce human intelligence. Yeah, creating machines that act like us is fascinating, but EA reasoning is not human, and if it were, it wouldn’t be very interesting,” Pierre Collet says. And the scientist doesn’t hesitate to quote the computer scientist Edsger Dijkstra, given the Turing Award in 1972, who wrote that “Alan Turning reflected on criteria to determine if machines can think,” but that “this question is as relevant as wondering if a submarine can swim.

Should we worry about a future of these kinds of programs and autonomous computers, capable of making decisions without us? “The next step in AI is to make machines inventive… not to replace the human, but to make him stronger, more creative, and to help him make decisions. The machine proposes, man decides,” Risto Miikkulainen says.

Deep learning improved by artificial evolution should have way more positive effects than negative, but that will always depend on what we make of these technologies, which are only tools,” Pierre Collet warns. At the ICube lab, the computer scientist supports setting up limits and safeguards, in particular by getting AI researchers to adopt an ethics charter, such as the one he signed in February 2017 along with 3,813 other scientists, including Risto Miikkulainen: the “23 principles of Asilomar.” Clearly inspired by the “three laws of Asimov,” it states that “the goal AI researchers aim for should be the creation of a beneficial AI, not an uncontrollable AI.” Our two scientists are staying vigilant.