Last week, while sipping my peach iced tea, I remember having an interesting conversation with this man. We spoke about the most bizarre thing you can imagine. Beer and diapers.
Doesn't make any sense, right? He told me about why supermarket stores started keeping baby diapers and beer next to each other, and the not-so-crazy reason behind it.
Reuters
What was the relation between beer and diapers? In the early 1990s, when people started analyzing shopping patterns with the help of computers, they found that a lot of guys go out to buy stuff on their own -- especially dads. When their wives sent them out on a re-supply run to the supermarket, these dads not only bought diapers for their kids but also threw in some beers in their shopping cart while they were at it. Mission accomplished.
"But in this example, at least there's some explanation," the conversation continued, with some chuckles. "A continuing challenge in machine learning and deep learning is how do you interpret results and understanding the ¡®why¡¯ behind a pattern or analysis."
I nodded my head at Wei Li, as he finished making his point. Dr Li is currently the Vice President and General Manager of Machine Learning and Translation at Intel. He's on his maiden visit to India, speaking at a technology conference. Dr Li tries to attend as many AI conferences as he can, he tells me.
"I feel like all this momentum around AI is the beginning of something big," he continues. "When I travel around the world, I see how different countries are trying to work on AI technologies, where everybody wants to do something."
reuters
Unless you've been living under a rock, you might've noticed how in the past year or so there¡¯s been a lot of marketing hype around AI, and Dr Li agrees. "Anybody can claim anything and before you know it everyone¡¯s an AI person," he says with a short laugh.?
Clearly AI is the buzzword of technology world right now, and there¡¯s a lot of interest and enthusiasm around the potential of artificial intelligence. But surely there must be some reality check, right?
"I'm not sure everyone understands the kind of problems AI can solve," Dr Li explains, "And when you don't know things one tends to go to extremes. Either the stuff is junk and useless, or the thing can solve all the problems you have. And the truth is somewhere in between."
Dr Li tells me how AI is just an umbrella term. How within AI there are things like machine learning, deep learning, visual computing, image recognition, natural language processing, and so on. Different AI technologies can solve different problems and not every problem can be solved today.
"So it does take some kind of understanding to know how AI can be useful and where the limitations are," says Dr Li. "You have to be positive about its potential, but also realistic, because when you think about today there¡¯s only certain things that you can do with AI."
Since AI, as a topic, has been in existence for over half a century, it¡¯s difficult to understand what stage of maturity it¡¯s at right now. "I think it¡¯s still very early days and AI technology is still evolving. The latest breakthrough on AI is on machine learning, and specifically on deep learning," Dr Li explains, "And image recognition is its most successful example, probably the most mature in this field."
But then there¡¯s also natural language processing, language translation and so on. People want to do predictive analysis in areas like medicine, for drug discovery, and all of these different areas have differing levels of AI maturity. And a lot of branches of AI still need data -- and lots of it -- for training and drawing meaningful conclusions.
So is there a need to ensure an AI system remains independent of human biases? "It¡¯s a big concern right now, a big issue," says Dr Li while sounding pretty grim. He illustrates this to me through an example of border control. Imagine an AI system that identifies ¡°bad people¡± based on their looks and skin tone alone, based on a heavily biased training phase that targets people of Asian descent -- that¡¯s a big problem. An example of AI being influenced by human bias.
"But the good thing is that people are aware of this problem," Dr Li says reassuringly. He hints at research happening on this very topic at university level. "However, there¡¯s no solution for this yet, no robust solution. Maybe someday we will have an AI free of human bias."
PwC
Dr Li paused before he uttered this next line, which I found quite fascinating. Bias in AI maybe bad, but we should have room for imperfection. "Machines don¡¯t necessarily have to be perfect," he says. As long as there¡¯s a standard wiggle room, we should be good.
Wait a minute, that doesn¡¯t make a lot of sense. Why build sophisticated AI or machines if they can¡¯t be perfect. Or at the very least be better than us humans? Dr Li responds simply, "We need to have a realistic expectation of what an AI or machine is supposed to do."
He further explained by referencing the self-driving car accident that killed a woman recently. ¡°With the first self-driving car accident, everybody got scared and started saying this may not work,¡± remembers Dr Li, ¡°but what we don¡¯t realize is that so many people get killed everyday on the road, and also so many people drive safely without killing anyone.¡± Doesn¡¯t mean we all suddenly stop driving cars. That would be an unfortunate knee jerk reaction on the path of AI progress, right?
To avoid unfortunate incidents and fully harness the power of AI, you need skilled people who can shape the future of AI and how it evolves. Unfortunately, Dr Li tells me, that¡¯s a problem right now.?
"I think there¡¯s a big skill gap that we need to fix quickly. Because as I attend conferences and talk to people, people who want to apply AI and are very interested, but don¡¯t necessarily have the expertise to do it."
Part of the solution, according to Dr Li, is to create more data scientists. "Wherever you go these days, we are talking about creating a data science program or data science department."
But that¡¯s not all. "The other part of the solution is to focus on how to make AI study and application easier," continues Dr Li. "Because if harnessing the power of AI means having a lot of PhDs in data science, then we¡¯ve failed as a community."?
Dr Li's vision for AI is something that would make it almost as a commodity. Just like assembling a car -- where somebody works on the shell, there are different experts for the tyres, engine, upholstery, paint, etc. -- where everyone has a key skill but works towards finishing a car. Access to AI and AI tools must be that easy for anyone to work on a solution.
The role of a data scientist, which is still in a lot of flux right now, will settle into someone like a solution architect. Someone who looks at the task at hand, breaks it down to different layers of programmers below him, understands statistics and Python, and tries to build AI-based solutions to accomplish certain tasks.
Obviously, our conversation veers towards an out-of-control AI -- how can any talk on AI be complete without the mention of Skynet. I mention Elon Musk and Stephen Hawking¡¯s warning on AI. Are people worried about nothing??
"I wouldn¡¯t say people worry about nothing," began Dr Li, "because if we aren¡¯t careful, bad things will happen because of AI. But we¡¯re too early worrying about things like Skynet versus getting benefits of AI, at this point." Anything we do involves risks and benefits, and at this point there¡¯s potential for a lot of AI benefits with too little risk.?
Between the Internet and AI, which one is a bigger disruption, I ask Dr Li. "AI has the potential to be a bigger disruption than the Internet," he tells me. We are still at the early stages of systems that can talk to you like a human, something that can do more than just exchange information between two terminals. A system that can learn and take independent decisions, in some limited capacity. "That's a game changer."
I swallow the last sip of iced tea from my glass, before asking Dr Li, a man who's on the cutting edge of AI research, working at a company that could have a big say in how mass-market AI solutions get deployed around the world, about what keeps him going and what keeps him up at night. Dr Li thinks for a few seconds before replying, "My biggest hope is that AI can be everywhere and help people in a positive way." He pauses. "And my biggest fear, for whatever reason, is if AI isn¡¯t used properly. If it isn¡¯t used well, things may get out of control. Because once it¡¯s out there, it¡¯s out there."
Wei Li is the Vice President and General Manager of Machine Learning and Translation at Intel. He received his Ph.D. in Computer Science from Cornell University, and also did his Executive Accelerator Program at the Stanford Graduate School of Business. These days he's working on the cutting edge of AI, developing software systems that include machine learning, binary translation, and hardware-software co-design.