Tech Trends Episode 9 Transcript
Anish: Welcome to TechTrends. TechTrends is a podcast series that provides prospective on the latest trends and technology, fintech and digital. Today’s episode is titled, “Embracing Artificial Intelligence,” and will provide insight into how artificial intelligence works, how it’s evolving, why people are so excited about it, and how it’s becoming an imperative for business to want to maintain a competitive edge. I’m Anish Bhimani, Chief Information Officer for Commercial Banking, and with me today is Apoorv Saxena, Global Head of Artificial Intelligence and Machine Learning for JPMorgan Chase. Apoorv, welcome to TechTrends.
Apoorv: Thank you, Anish. Great to be here.
Anish: Lot of excitement around artificial intelligence. It’s in the news every day, we hear a lot about it, and we also hear a lot about machine learning and deep learning. So, for people that are maybe less familiar with all those terms, can you talk about the difference?
Apoorv: Sure. Artificial intelligence is a pretty broad field of study. And its goal is trying to replicate human intelligence and exceed it over time. And machine learning is a soft field within artificial intelligence. This is more about than building software to recognize data patterns without explicitly programming it. And then deep learning is one of the techniques used in machine learning which is modeled more around how human brain works to make decisions and predictions, so on so forth.
Anish: So, we also hear about things like neural networks and supervised learning and unsupervised learning. Can you talk about some of the different types of machine learning?
Apoorv: Supervised learning is very similar to if I had the dog analogy: parent teaching the baby, “This is a dog. This is a cat.” So where you know the patterns that you want to recognize and feed the machine learning model or software with very specific details of how to recognize. Unsupervised learning is when a baby learns on its own, starts classifying and understanding the world. And in this case, the software understands and classifies, finds patterns without explicitly submitting telling. So that's how do you differentiate these two. And you have seen successes in both in terms of-- And those techniques, or types of machine learning, apply in different use cases.
Anish: Things have really taken off in the last 5 or 10 years. You see machine learning all around you, it's almost in everything we do, right? But why now? Why at this point?
Apoorv: Yeah, machine learning’s definitely has taken off in the last 10 years, it's almost every evolution happening. AI’s industry has gone through ups and downs and this is definitely an upcycle for sure. You must have heard about AI winters before. What is driving the latest sort of revolution is availability of data and cheap compute. And we need specialized hardware to do computing at scale. This is where old techniques, techniques that were developed in the 20th century, are being applied with new data. A great example of this is translation. The amount of accuracy improvement we’ve seen in translation is really driven by not new techniques, but huge amount of translated data available on the web. Same thing with speech recognition that is powering Alexa today. A lot of availability of transcribes audio. And so that's one set of drivers of why AI’s sort of going through evolution. Other set of drivers is a newer techniques being invented or a combination of newer techniques. So deep reinforcement learning is another example, where traditionally computer systems were not very good at predicting or making complex decisions, decisions that required two or three levels of thinking. Can see examples of that very complex software like AlphaGo, where computer software beat the most powerful player in AlphaGo. And same thing with video. Now your DOTA 5 is an example of open-source software that can beat humans in video gaming. So very powerful techniques being developed.
Anish: So there's a lot of excitement right now, right? It’s been said nobody knows what the question is, but the answer is definitely machine learning, right? Why is there so much excitement now?
Apoorv: There's definitely a lot of excitement, but you always have to temper that a little bit about some of the reality check. Clearly AI has been very successful and you're seeing AI everywhere. Alexa is a classic day-to-day example of highly specialized speech recognition being used to have a conversation with you, and my kid thinks Alexa is alive. The way Gmail helps you compose your email now. It's very powerful; it starts thinking about how you write a sentence versus anybody else. So you’re seeing application of AI everywhere and that gives you a sense of, “Hey the art of possibility, the curve of possibility keeps increasing. People start applying.” But it's very important to sort of draw an analogy to understand how complex this could be. So I go back to photography. It's an amazing piece of technology that you carry around on your phone but it took humans almost 200 years to develop all the way from a pinhole camera in early 1800s, to black and white photo, to color photo, to digital photo, to motion pictures, and now videos and photos. And all that photography does is essentially replicate the function of a human eye. And artificial intelligence as a field is trying to replicate the function of a human brain, which is an orders of magnitude more complex than human eye. AI has been around 60 or 70-year-old as a field, started by Alan Turing, as the father of AI. And so if you think about that, we are still very early.
Anish: It's interesting you use photography as an example. Yeah, it was 200 years for us to get there. But the last 20 or so have been exponential growth, right? I mean I think about 30 years ago, I was still developing film in a darkroom, right? And now my kids don't even know what film is. Is there an analogy there too? When do we hit that inflection point where you start to grow exponentially?
Apoorv: Clearly, I think the pace of innovation is accelerating. And the best reflection of that is quality of papers published in this industry, it has grown exponentially. And what gives me pause is essentially, there are some very fundamental unsolved problem in AI today. For example, deep learning, we still doesn’t have a very sound [theoretical?] underpinning. It’s an art much as a science…
Apoorv: …to get a really great results from your deep learning model.
Anish: If we are so early in that broad long-term journey, what are the use cases people are using today?
Apoorv: AI is very powerful and very useful already. Companies are using this to out-compete their competitors, and countries are now strategically thinking on AI policy for the whole country. And so, we are seeing that in retail, for example. AI is being used to do highly personalized shopping experience or personalized products, all the way to completely cashier-less checkout-free automated store. Same thing in healthcare. You're seeing AI being used to do better diagnosis than a doctors, or come up with new cures, or new drugs. And in our industry, in the financing overall, you're seeing AI being used to do better underwriting, improve how you do trading transaction or automated transactions. It’s being used to better detect fraud and so on so forth. The application of AI already here and real and being used widely every day.
Anish: So lots of use cases. How should companies think about incorporating artificial intelligence into their business?
Apoorv: I think the way you start off is: first you have to have the right talent to seed your sort of AI strategy. Seeding your team with the right talent from outside, or internally, is probably the first step. The second step you have to have is make this small team productive by having the right tools, the right data infrastructure, the right computer infrastructures. Unless you invest them, you will not be making them successful. And the third piece is, as I think, pick up the right use cases and taking that portfolio-based approach is critical. And the fourth piece I would say is take a much more holistic view of how AI’s going to be transformative for a business. AI’s going to be pervasive. So, you have to start, “How can I rethink my business processes? How do I run my business differently?” And if you don't have the right governance structure and the right business model, you will not fully utilize the power of AI.
Anish: When the internet still relatively new and companies were trying to figure out their online strategy and a lot of companies just sort of took the way they did business right now and put it online. But the companies that were really successful fundamentally transformed their business processes to take advantage of that.
Apoorv: You look at your call center serv-- customer service operation and say, “How can I use the latest speech recognition technology?” When a call comes in, I'm able to, in real-time, transcribe and suggest to the customer agent what are the things they should be speaking to their customer? That's transformative customer experience. And that’s how leading companies in this space are using AI to transform their business.
Anish: Yeah, and the call center example is a good one, right, because you also hear examples of people using voice recognition as authentication, right, using biometric authentication. And then so when you call in, you start speaking, I identify who you are and I pull up all of your data relevant to the call at the time, rather than a lot of the experience today where you have to jump through hoops. It's just a much better experience, right?
Apoorv: Exactly. You have to rethink which pieces of your business process or whole business process can be completely transformed. The example you gave of authentication using voice recognition. That was just not possible like even five years back. But now, if you are a company, you don’t have to invest a lot to get that from a third-party software can provide you that functionality into your sort of business process. And that's the power of how AI’s getting commoditized and being useful for everybody.
Anish: It's inevitable that whenever you start talking about automation, people start thinking about the impact on jobs, the impact on society. So how do you think about the responsibility of, whether it be technologists or technology companies, or companies leveraging artificial intelligence, to the workforce and where you should use AI and where you shouldn't and things like that?
Apoorv: I think it's important to distinguish between the short-term impact and the long-term impact. In the short-term, I believe AI will be more transformative in assisting human decision rather than completely automating and removing humans in the loop. In the long-term, I think it's probably fairly accurate to say a lot of the human jobs will be eliminated and it's irresponsible to not to acknowledge that. And as practitioners, as technologists, as business leaders, I think we have to collectively look at this. And this has happened many times in human history. New technology comes in, completely transforms workforce, and then the society’s role is to adjust to that. And something similar has to happen, but I think we have to be thoughtful of how fast that is done and how we sort of incorporate the transformation.
Anish: Whenever we talk about artificial intelligence and automation and stuff, people automatically start thinking about, you know, Skynet and the Terminator and the robots taking over, and other things like that. So how do we think about governance? How you manage artificial intelligence, where you should use it, where you shouldn’t? How do you think about machine learning models and making sure they work that they're supposed to and things like that?
Apoorv: I think you have to break it down because AI governance, it really comes down to three things. Number one, it comes down to data governance. You have to understand the data that you are building AI models on sometimes might be owned by you or might be owned by a customer. Do you have the right permission? Do they understand the intent on which you are creating your AI model? And I think miscommunication or misalignment there could cause serious problem, as we have seen in industries, in the sort of recent news around us. The second piece is even if you take the right-- if you have the right permission, you have the right governance model on data, the models that you build, where do you use them? Do you understand the biases that the model might have? Some biases are good. They're not necessarily bad. But where the model bias is not good for you or for your business is critical to understand. Having your governance structure around what is allowed, not allowed. So, I believe to have differentiation in a government structure around that is important. And the third piece is use case governance is what use cases, where should you and should you not apply AI. Great example I give is applying AI to filter resumes. Without thoughtful understanding of model bias, it’d be very detrimental to your business, potentially leave you up for potential lawsuits. But AI applied to target specific customer segment is perfectly legal. So I think those new answers of where AI is applied and what use cases you go after is going to be very critical as you go about building a government structure around AI in total.
Anish: So you hit on two important points in there. Number one is around model explainability, right, and being able to show why a model made a decision that it did or didn’t, right? But I think you also made an interesting point there is you have to be able to have explainability in certain cases, but you also have to have processes that are nimble enough to make sure that if you don't have those requirements, you can sort of move it along quickly. Is that fair?
Apoorv: Yes. Like any other new technology I think, unless the regulatory framework or the governance framework is not nimble, you will be imposing age-old models and processes developed for a different technology applied to same technology and stifle innovation. So I think we have to be very thoughtful about this, and at J.P. Morgan, of course we are spending a lot of time thinking about what is necessary and what is not necessary.
Anish: And then the second point you brought up was around bias, right, which is a huge topic of discussion right now. How much of the bias discussion is about making sure that the data you use to train a model is broad-based enough and diverse enough and other things like that versus the model itself?
Apoorv: I think it's both. I think a lot of the model biases originate from because the underlying data to train the model was biased. And the reason is the way collected this data brought in the human bias that is normal in society. So I think that's one piece to it. And then other piece is you can use very specific model techniques to remove bias. But again, you have to be very thoughtful around what model bias is good and what bias is not good for you as a business.
Anish: We've talked about the promise of artificial intelligence, we've talked about what can go wrong, how you put the right processes in place around that. Let's talk about how companies can get started with artificial intelligence. We started, sort of, for larger companies how they should think about this and how they should get started. You could talk a little bit about how we’re structuring our efforts around this.
Apoorv: Yeah, from larger companies, it's important to understand you have to be long-term thinking in terms of how transformative AI has to be. So, a lot of these large companies, JPMorgan or Google or Facebook, are looking at AI as transformative and all of these guys are thinking about 5-10-year strategy about AI. So the way you see-- do that is, at least at J.P. Morgan, is we started with seeding the company with talent, in some cases from outside, some cases in homegrown. We are bringing the best of the industry knowledge and then starting with that. Then seeding and creating an infrastructure in the business processes and compute layer to sort of enable and make these guys productive. I think that's the second piece. And I think the third piece companies like us are doing is taking a much more portfolio-based approach. Not only betting on something which is actionable and doable in the next 6 months to 11 months to 12 months, but also far-reaching. So we have hired Manuela Veloso who’s working on how you can apply AI for completely transforming the business in 5-6, 10 years from now. High-risk, but also potentially high-reward applications. They have hired people like me who are AI practitioners. As we adopt AI more generally, you will see a much more seeding of the talent within the businesses because that's where the transformation will happen. So that's the sort of approach we have taken at J.P. Morgan.
Anish: Yeah, and I think that’s an important point, which is you need to have a central team of experts that are building, whether they be data assets or models that can be used in a large enterprise because it's the only you can scale as to build those sort of platforms, right? But then I think it’s equally important to have people embedded in the business that understand the business and have enough appreciation of what artificial intelligence can do to actually drive that business transformation. It’s the right partnership that works well.
Apoorv: Exactly. As the organization learning matures and has the technology matures, I increasingly feel that these teams will dissolve into the businesses. Like you don't have a mobile strategy centralized team anymore. Mobile teams, digital teams are embedded closely with the business and I see the same evolution happening in AI or some of the newer technologies.
Anish: Okay, so that's larger companies, right? But for some of our listeners that might be at smaller firms maybe don't have the resources to build out their own central machine learning team, let alone embedded teams or data assets something like that, how should they think about getting started here?
Apoorv: So, I think two things they should know. One is they understand their business better than anybody else. So don't expect somebody from outside come in and telling you or consulting you how you apply AI into your business. So I think that’s number one. So, you have to own the business use cases, you have to own which use cases are alone for you. The second piece you have to understand is the barrier to entry to AI or deploying AI’s decreasing, increasingly. And as a result, you don't have to need a team of PhDs, or you don't need to spend multiple million dollars to get started. You can start looking at specific parts of your business process using third-party tool or open-source or some consultants. You can start in building something which gives you the confidence of how this technology works, builds organizational learning and the momentum needed to invest further. So I think start early, start small with a lot of third-party software, is probably where you will start. But take a long-term view. I think it's important, as you see success in your businesses, if you don't have a long-term view of this and how-- you will go back to the same thing we were talking about. When online boom happened, a lot of-- almost everybody created a dot com website. But unless you looked at it from a transformation perspective, internet allowed companies which invested in the long-term to be very successful and didn't went bankrupt.
Anish: So just to summarize, basically what you're saying is start small, pick a couple, maybe to use your example, low-risk use cases that are out there to get your feet wet, understand how this works, etcetera, maybe put some quick wins on the board, right? And then take a step back once you're more comfortable with the technology and sort of think about what are those transformational use cases you want to go after to really drive your business. Finally, Apoorv, what's next? So, when we have you back in a year for volume three of our machine learning conversations, right, or two years or five years, right, what’s coming down the pipe for artificial intelligence?
Apoorv: I think this, first of all, a great time to be in this space. So many things are happening, and across the board it's-- new techniques are being invented, number of research papers being published in this industry are exponential increasing. And that, basically what it means is newer techniques are being invented and which will be transformative five years from now. Some of the-- GAN is a very good example of a technique which was invented in 2014, which allow you to now make fake videos very easily. It's of course debatable how useful that is, but it's being used to do a lot more creative work. Now you can create very amazing videos with very small amount of compute power or talent, and that's transformative from a media perspective. In the medium-term, I see a lot of the technologies that have almost 99% solved, the last one person is still remaining. Self-driving car is a great example. AI-assisted driving is already here. Tesla is a great example. But I'm hoping to see self-driving car in the near future.
Anish: Self-driving cars always feel like they're perpetually two years away, right?
Anish: Apoorv, thanks very much for joining us today.
Apoorv: Thank you.
Anish: And I want to thank all of our listeners for joining us today. Tune in next time.