Why AI needs a broader, more realistic approach

THE CONCEPT of artificial intelligence (AI), or the ability of machines to perform tasks that typically require human-like understanding, has been around for more than 60 years. But the buzz around AI now is louder and shriller than ever. With the computing power of machines increasing exponentially and staggering amounts of data available, AI seems to be on the brink of revolutionizing various industries and, indeed, the way we lead our lives.

Vishal Sikka until last summer was the CEO of Infosys, an Indian information technology services firm, and before that a member of the executive board at SAP, a German software firm, where he led all products and drove innovation for the firm. India Today magazine named him among the top 50 most powerful Indians in 2017. Mr. Sikka is now working on his next venture exploring the breakthroughs that AI can bring and ways in which AI can help elevate humanity.

Mr. Sikka says he is passionate about building technology that amplifies human potential. He expects that the current wave of AI will “produce a tremendous number of applications and have a huge impact.” He also believes that this “hype cycle will die” and “make way for a more thoughtful, broader approach.”

In a conversation with Knowledge@Wharton, Mr. Sikka, who describes himself as a “lifelong student of AI,” discusses the current hype around AI, the bottlenecks it faces, and other nuances.

Knowledge@ Wharton: Artificial intelligence (AI) has been around for more than 60 years. Why has interest in the field picked up in the last few years?

Vishal Sikka: I have been a lifelong student of AI. I met [AI pioneer and cognitive scientist] Marvin Minsky when I was about 20 years old. I’ve been studying this field ever since. I did my Ph.D. in AI. John McCarthy, the father of AI, was the head of my qualifying exam committee.

The field of AI goes back to 1956 when John, Marvin, Allen Newell, Herbert Simon and a few others organized a summer workshop at Dartmouth. John came up with the name “AI” and Marvin gave its first definition. Over the first 50 years, there were hills and valleys in the AI journey. The progress was multifaceted. It was multidimensional. Marvin wrote a wonderful book in 1986 called The Society of Mind. What has happened in the last 10 years, especially since 2012, is that there has been a tremendous interest in one particular set of techniques. These are based on what are called “deep neural networks.”

Neural networks themselves have been around for a long time. In fact, Marvin’s thesis was on a part of neural networks in the early 1950s. But in the last 20 years or so, these neural networkbased techniques have become extraordinarily popular and powerful for a couple of reasons.

First, if I can step back for a second, the idea of neural networks is that you create a network that resembles the human or the biological neural networks. This idea has been around for more than 70 years. However, in 1986 a breakthrough happened thanks to a professor in Canada, Geoff Hinton. His technique of backpropagation (a supervised learning method used to train neural networks by adjusting the weights and the biases of each neuron) created a lot of excitement, and a great book, Parallel Distributed Processing, by David Rumelhart and James McClelland, together with Hinton, moved the field of neural Net-related “connectionist” AI forward. But still, back then, AI was quite multifaceted.

Second, in the last five years, one of Hinton’s groups invented a technique called “deep learning” or “deep neural networks.” There isn’t anything particularly deep about it other than the fact that the networks have many layers, and they are massive. This has happened because of two things. One, computers have become extraordinarily powerful. With Moore’s law, every two years, more or less, we have seen doubling of price performance in computing. Those effects are becoming dramatic and much more visible now. Computers today are tens of thousands of times more powerful than they were when I first worked on neural networks in the early 1990s.

The second thing is that big cloud companies like Google, Facebook, Alibaba, Baidu and others have massive amounts of data, absolutely staggering amounts of data, that they can use to train neural networks. The combination of deep learning, together with these two phenomena, has created this new hype cycle, this new interest in AI.

But AI has seen many hype cycles over the last six decades. This time around, there is a lot of excitement, but the progress is still very narrow and asymmetric. It’s not multifaceted. My feeling is that this hype cycle will produce great applications and have a big impact and wonderful things will be done. But this hype cycle will die and a few years later another hype cycle will come along, and then we’ll have more breakthroughs around broader kinds of AI and more general approaches. The hype we see around AI today will pass and make way for a more thoughtful and realistic approach. Knowledge@ Wharton: What do you see as the most significant breakthroughs in AI? How far along are we in AI development?

Sikka: If you look at the success of deep neural networks or of reinforcement learning, we have produced some amazing applications. My friend [ and computer science professor] Stuart Russell characterizes these as “one- second tasks.” These are tasks that people can perform in one second. For instance, identifying a cat in an image, checking if there’s an obstacle on the road, confirming if the information in a credit or loan application is correct, and so on.

With the advances in techniques — the neural networkbased techniques, the reinforcement learning techniques — as well as the advances in computing and the availability of large amounts of data, computers can already do many one- second tasks better than people. We get alarmed by this because AI systems are superseding human behavior even in sophisticated jobs like radiology or legal — jobs

that we typically associate with large amounts of human training. But I don’t see it as alarming at all. It will have an impact in different ways on the work force, but I see that as a kind of great awakening.

But, to answer your question, we already have the ability to apply these techniques and build applications where a system can learn to conduct tasks in a welldefined domain. When you think about the enterprise in the business world, these applications will have tremendous impact and value.

Sikka: You find fraud by connecting the dots across many dimensions. Already we can build systems that can identify fraud far better than people by themselves can. Depending on the risk tolerance of the enterprise, these systems can either assist senior people whose judgment ultimately prevails, or, the systems just take over the task. Either way, fraud detection is a great example of the kinds of things that we can do with reinforcement learning, with deep neural networks, and so on.

Another example is anything that requires visual identification. For instance, looking at pictures and identifying damages, or identifying intrusions. In the medical domain, it could be looking at radiology, looking at skin cancer identifications, things like that. There are some amazing examples of systems that have done way better than people at many of these tasks. Other examples include security surveillance, or analyzing damage for insurance companies, or conducting specific tasks like processing loans, job applications or account openings. All these are areas where we can apply these techniques. Of course, these applications still have to be built. We are in the early stages of building these kinds of applications, but the technology is already there, in these narrow domains, to have a great impact.

Sikka: It is human nature to continue what has worked, so lots of money is flowing into ongoing aspects of AI. From chips, in addition to NVidia, Intel, Qualcomm etc., Google, Huawei and others are building their own AI processors and many start-ups are as well, and all this is becoming available in cloud platforms. There are tons of work happening in incrementally advancing the core software technologies that sit on top of this infrastructure, like TensorFlow, Caffe, etc., which are still in the early stages of maturity. And this will of course continue.

But beyond this, my sense is that there are going to be three different fronts of development. One will be in building applications of these technologies. There is going to be a massive set of opportunities around bringing different applications in different domains to the businesses and to consumers, to help improve things. We are still woefully early on this front. That is going to be one big thing that will happen in the next five to 10 years. We will see applications in all kinds of areas, and there will be applicationoriented breakthroughs.

Two, from a technology perspective, there will be a realization that while the technology that we have currently is exciting, there is still a long way to go in building more sophisticated behavior, building more general behavior. We are nowhere close to building what Marvin [Minsky] called the “society of mind.” In 1991, he said in a paper that these symbolic techniques will come together with the connectionist techniques, and we would see the benefits of both. That has not happened yet.

John [McCarthy] used to say that machine learning systems should understand the reality behind the appearance, not just the appearance. I expect that more general kinds of techniques will be developed and we will see progress towards more ensemble approaches, broader, more resilient, more general- purpose approaches. My own Ph. D. thesis was along these lines, on integrating many specialists/narrow experts into a symbolic generalpurpose reasoning system. I am thinking about and working on these ideas and am very excited about it.

The third area — and I wish that there is more progress on this front — is a broader awareness, broader education around AI. I see that as a tremendous challenge facing us. The development of AI is asymmetric. A few companies have disproportionate access to data and to the AI experts. There is just a massive amount of hype, myth and noise around AI. We need to broaden the base, to bring the awareness of AI and the awareness of technology to large numbers of people. This is a problem of scaling the educational infrastructure.

Knowledge@ Wharton: Picking up on what you said about AI development being asymmetric, which industries do you think are best positioned for AI adoption over the next decade?

Sikka: Manufacturing is an obvious example because of the great advances in robotics, in advancing how robots perceive their environments, reason about these, and affect increasingly finer control over it. There is going to be a great amount of progress in anything that involves transportation, though I don’t think we are still close to autonomy in driving because there are some structural problems that have to be solved.

Health care is going to be transformed because of AI, both the practice of health care as well as the quality of health care, the way we build medicines, proteinbinding is a great case for deep learning, personalize medicines, personalization of care, and so on. There will be tremendous improvement in financial services, where in addition to AI, decentralized/p2p technologies like blockchain will have a huge impact. Education, as an industry, will go through another round of significant change.

There are many industries that will go through a massive transformation because of AI. In any business there will be areas where AI will help to renew the existing business, improve efficiency, improve productivity, dramatically improve agility and the speed at which we can conduct our business, connect the dots, and so forth. But there will also be opportunities around completely new breakthrough technologies that are possible because of these applications — things that we currently can’t foresee.

The point about asymmetry is a broader issue; the fact that a relatively small number of companies have access to the relatively small talent of people and to massive amounts of data and computing, and therefore, development of AI is very disproportionate. I think that is something that needs to be addressed seriously.

Sikka: I find it extraordinary that in the traditional industries, for example in construction, you can walk into any building and see the plans of that building, see how the building is constructed and what the structure is like. If there is a problem, if something goes wrong in a building, we know exactly how to diagnose it, how to identify what went wrong. It’s the same with airplanes, with cars, with most complex systems.

But when it comes to AI, when it comes to software systems, we are woefully behind. I find it astounding that we have extremely critical and extremely important services in our lives where we seem to be okay with not being able to tell what happened when the service fails or betrays our trust in some way. This is something that has to be fixed. The compartmentalization of data and broader access to it has to be fixed. This is something that the government will have to step in and address. The European governments are further ahead on this than other countries. I was surprised to see that the EU’s decision on demanding explainability of AI systems has seen some resistance, including here in the valley.

I think it behooves us to improve the state of the art, develop better technologies, more articulate technologies, and even look back on history to see work that has already been done, to see how we can build explainable and articulate AI, make technology work together with people, to share contexts and information between machines and people, to enable a great synthesis, and not impenetrable black boxes.

But the point on accessibility goes beyond this. There simply aren’t enough people who know these techniques. China’s Tencent sponsored some research recently which showed that there are basically some 300,000 machine learning engineers worldwide, whereas millions are needed. And how are we addressing this? Of course there is good work going on in online education and classes on Udacity, Coursera, and others. My friend [ Udacity cofounder] Sebastian Thrun started a wonderful class on autonomous driving that has thousands of students. But it is not nearly enough.

And so the big tech companies are building “AutoML” tools, or machine learning for machine learning, to make the underlying techniques more accessible. But we have to see that in doing so, we don’t make them even more opaque to people. Simplifying the use of systems should lead to more tinkering, more making and experimentation. Marvin [ Minsky] used to say that we don’t really learn something until we’ve learnt it in more than one way. I think we need to do much more on both making the technology easier to access, so more people have access to it, and we demystify it, but also in making the systems built with these technologies more articulate and more transparent.

Sikka: As I mentioned earlier, research and availability of talent is still quite lopsided. But there is another way in which the current state of AI is lopsided or bottlenecked. If you look at the way our brains are constructed, they are highly resilient. We are not only fraud identification machines. We are not only obstacle detection and avoidance machines. We are much broader machines. I can have this conversation with you while also driving a car and thinking about what I have to do next and whether I’m feeling thirsty or not, and so forth.

This requires certain fundamental breakthroughs that still have not been happened. The state of AI today is such that there is a gold rush around a particular set of techniques. We need to develop some of the more broadbased, more general techniques as well, more ensemble techniques, which bring in reasoning, articulation, etc.

For example, if you go to Google or [ Amazon’s virtual