Researcher wants respect for pioneering link to AI
1990s work is crucial to progress in speech and vision software.
LUGANO, SWITZERLAND — Jurgen Schmidhuber may be the Rodney Dangerfield o f a r t i f i c i a l i n t e l l i g e n c e research.
In a visit with him in this idyllic Swiss city in the mountains near the Italian border, it is easy to understand why he believes that his pioneering work in the field often, as the comedian liked to say, gets no respect.
Far away in Silicon Valley, on the other side of the world, the tech industry is building cars that drive themselves and household appliances that respond to your voice commands and even try to predict what you will do next.
In certain circles, the people who did the early work that made this technology possible are stars. There is Sebastian Thrun, a roboticist who did groundbreaking research on self-driving cars at Google. Adam Cheyer and Tom Gruber worked on the AI program Siri, later acquired by Apple. And Facebook hired Yann LeCun, an expert in “neural networks” who left New York University to start a research program at the social media giant.
But mention the name Jurgen Schmidhuber in an automated quinoa lunch spot frequented by coders in San Francisco, and you are likely to get blank stares.
On a recent train ride to Zurich, Schmidhuber, an a t hl e t i c 5 3 -ye a r- ol d who is co-director of the Dalle Molle Institute for Artificial Intelligence Research here, reflected on how he believed his early research was often overlooked or ignored. “It’s like much of the rest of society,” he said. “Sometimes it’s postfactual.”
Schmidhuber’s complaints are well known within the fraternity of researchers who have turned what until a half-decade ago was an academic backwater into a multibillion-dollar industry. He has been accused of taking c re di t for ot her people’s research and even using multiple aliases on Wikipedia to make it look as if people are agreeing with his posts.
“J u r g e n i s m a n i c a l l y obsessed with recognition and keeps claiming credit he doesn’t deserve for many, many things,” LeCun said in an email. “It causes him to systematically stand up at the end of every talk and claim credit for what was just presented, generally not in a justified manner.”
Schmidhuber counters that criticism with a bigger point: He is not the only one who is not getting due credit among AI researchers. In fact, he says work going all the way back to the 1960s is regularly ignored by today’s research luminaries.
Although he insists he does not harbor ill will toward those better-known researchers, it grates on him that history has not been kinder. “Certain researchers in my field have acted as if they invented something, although it was invented by other people whom they did not even mention,” Schmidhuber said.
But understanding the disconnect between his early work and his lack of celebrity is not easy — and cannot be entirely explained by the fact that he lives thousands of miles from the tech industry’s center of gravity.
The dispute is about the roots of neural net works, which allow machines to learn by recognizing patterns that can then be applied generally. Applications include r e c o g n i z i n g s p e e c h a n d language, vi sually identi - fying objects, navigating in self-driving cars and making robot hands grasp more deftly. As a scientific field, it dates to the 1940s. But only in recent years have research- ers in this area made striking progress.
Neural networks are actually soft ware. For a visual analogy, think of them as a giant Tinkertoy set — vast a r r ays of i nterc onnec te d nodes that can be trained to do everything from language translation to recognizing visual objects or human speech.
For decades, neural networks were laboratory curiosities, often met with skepticism. But in the 1990s, with faster and cheaper computers as well as new ideas about how to design neural nets, there was finally progress.
I n 1 9 9 7, S c h mi d h u b e r and Sepp Hochreiter publi shed a paper on a technique that has proved crucial in laying groundwork for the rapid progress that has been made recently in vision and speech. The idea, known as Long Short-Term Memory, or LSTM, was not widely understood when it was introduced. It essentially offered a form of memory or context to neural networks.
Ju s t a s humans d o n o t restart learning from scratch every second, a certain type of neural network adds loops or memory that interpret each new word or observation in light of what has been previously observed. LSTM strikingly improved these networks, leading to huge jumps in accuracy.
It may be that Schmidhuber’s misfortune is that he was simply too early — a few years ahead of the powerful and more affordable computers we have today. It was not until recently that his concepts started to pan out.
L ast year, for example, Google researchers reported that they had used LSTM to cut transcription errors in their speech recognition service by up to 49 percent. It was a huge increase after years of incremental progress.
B u t b e t we e n S c h mi d - huber’s and Hochreiter ’s research and today’s progress there was a big gap — and that is the rub. Other researchers say it took many contributors to get from Point A to Point B, where we are today.
“He’s done a lot of seminal stuff,” said Gary Bradski, an AI scientist who created a popular computer vision system known as OpenCV. “But he wasn’t the one who made it popular. It’s kind of like the Vikings discovering America; Columbus made it real.”
Schmidhuber also has a grand vision for AI — that s e l f- aware or “c onsc i ous machines” are just around the corner — that causes eyes to roll among some of his peers. To put a fine point on the debate: Is artificial intelligence an engineering discipline, or a godlike field on the cusp of creating a new superintelligent species?
Schmidhuber is firmly in the god camp. He maintains that the basic concepts for such technologies exist and that there is nothing magical about human consciousness.
“Generally speaking, consciousness and self-awareness are overrated,” he said, arguing that machine consciousness will emerge from more powerful computers and software algorithms much like those he has designed.
It has been an obsession since he was a teenager in Germany reading science fiction.
“As I grew up I kept asking myself, ‘What’s the maximum impact I could have?’” Schmidhuber recalled. “And it became clear to me that it’s to build something smarter than myself, which will build something even smarter, et cetera, et cetera, and eventually colonize and transform the universe, and make it intelligent.”
Today, he will not be pinned down on when such thinking machines might arrive, saying only that given the vast improvements in computing power it will be soon.
I n 2 014, he a nd ot her s founded a company to commercialize some of the technology that he helped create and to work on “general purpose” artificial intelligence.
The company, Nnaisense, is based just a few steps from the University of Lugano campus. It is being advised by Hochreiter, who now heads the Institute of Bioinformatics at the Johannes Kepler University in Linz, Austria, and Jaan Tallinn, a co-founder of Skype. The company has partnerships in finance, autonomous vehicles and heavy industry.
Nnaisense’s chief executive is an American computer scientist, Faustino Gomez, who has been Schmidhuber’s research collaborator for many years. He defends both his partner’s claims of having done pioneering work and his optimism about the field that has begun shaking up industries and economies around the world.
“We are at the beginning of the end of the beginning in AI,” he said.