OVER the past decade, machine intelligence has vastly improved. That is in large part due to deep learning, a technique that gives computers the ability to teach themselves. It underpins everything from world-beating chess and Go algorithms to digital voice assistants like Amazon鈥檚 Alexa and Apple鈥檚 Siri.
Yoshua Bengio is one of the pioneers of deep learning, and has spent his career at the forefront of AI research. He was recently awarded the A. M. Turing Award, which is often called the Nobel prize of computing, along with two other deep learning pioneers: Geoffrey Hinton at the University of Toronto and Google, and Yann LeCun, who is chief AI scientist at Facebook. The trio will split the $1 million prize.
Advertisement
Bengio remains in academia at the University of Montreal, Canada, but co-founded an AI incubator and advises on a couple of startups. He resisted the draw of a juicy Silicon Valley salary, because he believes 鈥渉umans are more important than money鈥�. That being said, he isn鈥檛 humanity鈥檚 biggest fan. For all his optimism about the future of machines, he wouldn鈥檛 put it past us to mess things up.
How do you think AI will be able to actually help people?
The progress we have made in machine learning has been pretty amazing, and it can empower almost any sector of society.
I have invested a lot of effort in healthcare applications, especially the goal of machines being able to diagnose cancer from medical images. I also think there is a lot of potential when it comes to climate change, such as using AI to help predict the consequences of societal transformations. And we have been very active in looking at how machine learning can help promote human rights, by detecting gender bias, for example.
Does AI have any potential risks?
How much time do I have? Lethal autonomous weapons, for example, are morally wrong because computers don鈥檛 understand moral values. That means they might not be able to question the legitimacy of an order. They are also wrong for security reasons, as they can threaten the global balance of power. That could be dangerous for all of us.
AI can also be used to control people, monitoring where they are and reinforcing the power of authoritarian governments in a way that wasn鈥檛 possible before. For example, we know that China has hundreds of millions of cameras in the streets, equipped with the technology for facial recognition.
鈥淏rains are very complicated machines, but I think we will figure out the principle of intelligence鈥�
Another related danger that people talk about less is how AI could be used to influence others. Think about advertising, the influence of which is usually seen as benign, and extend those techniques to politics. It might not matter if I buy Coke or Pepsi, but it does matter if I vote for Trump or not, right?
Then you have the issue of bias and diversity, where AI can reinforce some of the negative aspects of our current societies. And finally, there is the concentration of power. As a powerful technology left to its own devices, it is just going to reinforce the power of those who control it, and that is bad for democracy as well.
Do you think the big technology companies, like Google or Facebook, have too much power?
Yes. I think the trend is concerning. There is a snowball effect. The more data you have, the more customers you have. And if you are rich, you can pay for the best researchers. This is bad for innovation and it鈥檚 bad for democracy. It is bad for innovation, because innovation comes from diversity, from many different people with different goals trying different things. And then for democracy鈥� well, democracy, what does it mean? It is power to the people. If power is concentrated in a few hands, that isn鈥檛 good.
How do you feel about the future of AI?
I am very optimistic about the science making a lot of progress in the coming decades. I think brains are very complicated machines, but I think we will figure out the principles of intelligence, which will help us make better AI. So I am very optimistic on that side, but I also realise that it might take decades, or even centuries, for all we know. But we will get there, unless we self-destruct in some way as a social organisation.
Is an AI apocalypse something people should be afraid of?
Well I鈥檓 not.
Why not?
Because that scenario just doesn鈥檛 fit my understanding of the science of AI right now. I don鈥檛 see it as credible. Now, I don鈥檛 have a crystal ball, and the science of AI 50 years from now will be very different. I think we need to be prudent, and it is good that there are people who are thinking about these issues, but making it a political or social question at this point is very premature. We should be worrying about those other, shorter term issues that, for sure, are happening and need our attention.
Are the short-term issues unstoppable, or is there a way to halt them?
There is a way, but it isn鈥檛 an easy one. We need to have society at large understand those issues and bring them to the forefront of the political agenda, so that governments act as they should.
Finally, what do people get wrong about AI?
So many people overestimate the intelligence of these systems. AIs are really dumb. They don鈥檛 understand the world. They don鈥檛 understand humans. They don鈥檛 even have the intelligence of a 6-month-old. Yes, they can beat the world champion at Go. But that doesn鈥檛 mean they can do anything else.
A related misconception is that intelligent robots are taking over the world. People project their own emotions and feelings onto the machines that we will build in the future and think, 鈥渨ell, if I was that machine, I would be angry that all these guys have been enslaving me, and now I鈥檓 out for revenge鈥�.
But I think this is nonsense. We are designing those machines, which means the real danger is if an AI gets into the wrong hands, and is then used in ways that will hurt us. It isn鈥檛 that the AI is malevolent, it is the humans that are stupid and/or greedy.
