Jürgen Schmidhuber is painting an image of the future of our Universe. And it’s plain to see, we are neither a real part of it, nor is it our Universe at all.
“In 2050 there will be trillions of self-replicating robot factories on the asteroid belt,” he tells the audience at WIRED2016. “A few million years later, AI will colonise the galaxy. Humans are not going to play a big role there, but that’s ok. We should be proud of being part of a grand process that transcends humankind more than the industrial revolution. It is comparable to the invention of life itself, and I am privileged to live this moment and witness the beginnings of this.”
The pioneer in deep learning neural networks should know. His work - including 333 peer-reviewed papers - has formed the foundations of many of the AI systems we see embedded in smartphones today, including Google’s voice recognition and Google Translate - “each of you has a little piece of us in your pocket,” he said.
Although he sees a not-so-distant future where humans become fairly redundant, the AI that takes over would obviously not have been possible without us. Schmidhuber has been working on recurrent artificial neural networks for decades, and these systems are modelled after the human brain. They are an attempt to mimic the billions of neural connections firing away every second to deploy countless biological tasks and cognitive processes. AIs modelled on this are essentially learning algorithms: “Networks figure out over time which inputs are important and which aren’t.”
An outcome of this research was something Schmidhuber called ‘long-short-term memory’. “No-one was really interested at first - now all major companies are using it for speech recognition. It’s a universal mechanism. Google is using that same LSTM for speech recognition and translation. You train it on lots of examples.”
The first robots that deployed this kind of approach are “artificially curious systems”, says Schmidhuber. These robots “set goals like a baby setting new experiments to play around. Curiosity and creativity, coming up with new experiments so they can learn more about how the world works - this makes it so they become better problem solvers and improve their skills repertoire.”
Read more: Machine learning versus AI: what's the difference?
Much of the groundwork for these types of AI was first explored decades earlier, Schmidhuber concedes. We simply now have the cheap, fast computing power needed to make the processes work as they were first conceived: “In the previous millennium it took a while until computers were so fast that this method could find its potential. There was a dramatic breakthrough and they were twice as good as before.”
“Everyone thinks Google invented self-driving cars. But Ernst Dickmanns already had fast self-driving cars with cameras [in the 80s and 90s] that drove up to 100km/h.”
Today’s neural networks are able to carry out medical diagnoses - Schmidhuber points to his own team’s 2012 work on recognising pre-cancerous cells in the breast. “We train a stupid neural network to do the same as a doctor, based on lots of training examples. It becomes as good or better than the best competitor and rivals human performance now.” As a result of these types of projects, the likes of IBM and Google are exploring how AI can disrupt the medical sector, and startups are joining in.
But what lies between now and our total redundancy as a species, when AI starts exploring space for more resources and leaving us behind? Life as the planet’s overlords will get a bit better before AI pushes us out to setup its galactic factories. Schmidhuber adds that a few years from now we will see animal-like AIs that exhibit intelligence on a par with crows and capuchin monkeys. “Once we have that, it’s a small step to human-AI. It took just a few million years [of evolution] to get to human intelligence, and computers are faster.”
“Things are going to change dramatically in the near future.”
This article was originally published by WIRED UK
