- Davos 2016 - The State of Artificial Intelligence
Click on text below to jump to specific point in the video
Welcome to this session. It is a collaboration between World Economic Forum and Arirang TV. I am Jennifer Moon. AI is no longer science fiction. We will discuss its state today. Panelist introduction
Before we delve in, for our viewers we prepared a brief video. Professor Moore: Where are we today in 2016?2015 was a great year for AI: machine learning on a large scale, previously available only to big companies, is now available to researchers. Many people are leaving the big companies to create new startups in this field. Big lessons: emotional understanding, gradual work to remove the boring parts of white collar work. For example, in legal world, the boring parts of understanding legal documents is being taken away. One particular thing: lots of humans are involved in negotiating with other humans. In 2015 computers learnt to negotiate.
A couple of other things: 2015 you could take the hands off the wheel of a self driving car. Progress on understanding language in the legal area has been very important. Suppose search engines really start understanding, factually, all documents they process. The search industry could go from 1 trillion to 10 trillion dollars. Another consequence of the ability to understand language: the ideal personal assistant. A very capable personal assistant could be incredibly valuable for people with fewer economic resources. Ya Qin: How is Baidu adopting AI in industry? How do you plan to monetize it?AI is becoming mainstream. For Baidu it is essentially embedded in every product. We have a platform open to all teams within Baidu. We have now one of the world's largest deep neural nets. Having Open AI available to researchers. Qualcomm, you demoed some impressive products at CES last monthWe are seeing these technologies move out of the data center and into the world. Its the very widespread nature of devices which can now use these technologies that we are very excited about. Should I throw out my driver's license now?It will probably be a few years before you can throw away your license. Right now, Tesla's hands free mode only works on the highway. The reason: although perception is quite capable, the decisions about what to do is still a good old-fashioned rule based system. Every so often, the rules don't apply - for e.g. cyclist coming down the wrong way. The Google Car gets confused in that situation. A different approach involves not just rules, but also reasons why. To deal with unexpected circumstances, you need to endow the machine with knowledge. This kind of decision making technology will happen in a few years. Matt: What can an AI smartphone not do for me at this point, due to limitations in device capabilities?You really need the combination of algorithms and computing capability with a database. On a device, you have constraints on both, but with cloud computing both of those (processing and data) become available.We are trying to improve the algorithms as well as the capability of the hardware. Volume of database is probably something Baidu doesn't have much problem withLet me get back to self-driving. We just completed a test in Beijing. It will take years to be commercialized. We need high precision mapping. We need accurate positioning. It needs a lot of investment in infrastructure. But it will probably come faster than most people think. Why? The theme of the fourth industrial revolution is how fast it (and AI) is evolving. Will machines become smarter than humans?One by one we will see things which require personal ingenuity become automated. Many professions which we thought were smart - lawyers, doctors etc. - will diminish. AI will help humans in charge even where deep social interaction is required. We have assumed AI being an aid, and not a threat. What is it humans are worried about then?This is a longer term question. DeepMind demonstrated a learning system which in some ways resembles a new born baby. It has absolutely no preprogramming of any kind. It is given the screen of an Atari game - with no notion of time, speed, death etc. It is able to play at a human level in a few hours. This is a nice demonstration of generality. Also, these large learning networks are inscrutable - we don't really know what they are doing.
We know that those techniques don't extend to the wide range of tasks that humans do. But it might only be a small number of breakthroughs between now and general purpose computing. But they are very hard to predict.
I would argue that the possible risks are not immediate, but we need to keep them under control. We should do the research to take care of it starting now.
We don't know how to specify objectives very well, like Midas. The machine is going to carry it out. You might in effect be setting up a chess match between the machines and the human race, and we already know how that turned out.
I run a large AI university, and the faculty and students often worry about this issue. But they see ways to save lives now. In the back of our minds, we are very concerned about the safety of the autonomous systems. With narrow AI, we focus on making the systems safe and making sure they do what we say they are going to do. But the students and faculty right now are only rolling up their sleeves to use AI to save lives. Business and industry leaders: do you agree with this?We need to have caution. I don't think we need to worry about in the extremely near future. There is a lot more positive potential.
But we have to be mindful of security. But one can use AI to improve security. The applications are profound in the near term.
The community needs to be concerned of the overall direction. But in the short term, industry is investing in narrow AI.
However, as an aside, I am concerned about whether the human becomes too dependent on the machine and might become less intelligent. We might become lazy in thinking.
Another concern is social behavior change. As machine becomes more autonomous, we need to adapt to that. We know where AI is now. What would be the next game changer in AI?It is embarrassing to roboticists that we are really good at vision now, but we haven't progressed on picking objects. This will help for object manipulations with our exoskeletons. I want to clarify something. Is AI and the act of picking things up different?It just turns out that the things we thought were fancy and clever, turn out to be quite easy to implement. And other things, like picking up things, are very hard.
Its a chicken and egg problem. To be as dexterous as humans, the robot needs to have really really complicated hands. Its very expensive to develop that technology. Its possible that 3D printing might provide a breakthrough here, because we can quickly test and iterate on much more complex devices more cheaply. Applications in agriculture and elder care become feasible when robots can pick up objects. Matt: What does Qualcomm focusing on in terms of the next game changer in AI?We look at mobile use cases. Apps for smartphones - looking at your images and acting upon them, or looking at your context and perform a personal assistant type function effectively. Its a hard problem with a lot of subtlety. In robotics - medical devices to do diagnostics, self driving vehicles.
We are looking at search and interfacing with personal assistant. Technologies that can be applied to finance and healthcare. Questions from the audienceHow do you think AI will improve us? How about superhumans?I am not optimistic about uploading our mind to machine and living forever. We have no scientific theory of consciousness. But it is possible to help someone paralyzed to use robots to perform actions using their mind to control it. The amazing thing is - we don't understand which signals are used. This can overcome one of the greatest bottlenecks in human cognition - our short term memory. If we can multiply our short term memory by a factor of 10, we could dramatically improve human cognition. Question for all: technology is used to promote humanness. What would you like AI to achieve in terms of governance?One really important solution is to make sure all kids, especially girls, are encouraged into this area.
As for helping with education of youngsters, we have been seeing in the last few years that emotion understanding really helps with overall outcomes. Is AI working towards cyber-defense?There are really two applications in warfare: cyberwarfare and autonomous weapons.
Autonomous weapons - the machine chooses its own targets. With autonomous weapons, it is easy for non-state actors to create weapons of mass destruction to catastrophic effect.
Cyberwarfare is already going on. Its only a matter of time before cyber-warfare could trigger a real physical war. Wrap up thoughts from panelists?We are in a really exciting time. AI is being used for good across the planet. I encourage youngsters to get into this area.
AI is certainly the foundation, the engine and driver for the next decade's technologies. Consider AI if you start a business.
I couldn't agree more. Pragmatic advancements - those are all upon us.
Everything good we have is the result of our intelligence. If AI can amplify and provide tools for our intelligence, we could be talking about a golden age for humanity. I am optimistic the upside is great.
Thank you.Video outline created using VideoJots