Alyssa Simpson Rochwerger and Wilson Pang: Real World AI: A Practical Guide for Responsible Machine Learning

Transcript

[0:00:25] DA: When AI works, it’s nothing short of brilliant, helping companies make or save tremendous amounts of money while delighting customers on an unprecedented scale but when it fails, the results can be devastating. In their new book Real World AI, Alyssa Simpson Rochwerger and Wilson Pang share dozens of AI stories from startups and global enterprises alike, featuring personal experiences from people who have worked on global AI deployments that impact billions of people every day. AI for business doesn’t have to be overwhelming and their new book uses plain language to walk you through an AI approach that you could feel confident about for your business and for your customers. Hey listeners, my name is Drew Applebaum and I’m excited to be here today with Alyssa Simpson Rochwerger and Wilson Pang, author of Real World AI: A Practical Guide for Responsible Machine Learning. Alyssa, Wilson, thank you for joining, welcome to the Author Hour Podcast.

[0:01:17] ASR: Thanks for having us.

[0:01:18] WP: Thank you, it’s great to be here.

[0:01:21] DA: Let’s kick this off, can you give us a quick rundown of your respective professional backgrounds?

[0:01:28] ASR: Yeah, I have a little more than 10 years of product management experience, perhaps an unlikely expert in artificial intelligence and machine learning. I got into field while I was working at IBM, in the Watson division but I really became passionate about it because this technology is one that really can transform so many industries and really create new experiences that can be used to build products that people love and engage customers in new ways and solve problems. Currently, I’m knee-deep in problems of healthcare and trying to use data and technology and a little bit of machine learning to solve problems in the healthcare space.

[0:02:13] WP: Yeah, I’ve been with the technology industry for almost 20 years now and also got lucky, very lucky to start work on machine learning about 11 or 10 years ago when I was with eBay. We started the search science team and that is probably the first team who is using machine learning to optimize all those user experience. Then, got firsthand experience how machine learning kind of how to improve the overall e-commerce experience, how then it can help to increase their coercion and have the company to drive the revenue growth. Then I got a chance to work on many different other areas that like marketing, like operation, et cetera, all using machine learning. I have seen the power of machine learning and what machine learning can do and how machine learning can help. Then I joined Safe Trip, which is second biggest online travel company in the world, as their chief data officer where I oversee all the data analytics and machine learning for the whole group. My team also use machine learning a lot to transform the travel industry using machine learning to do CRM, using machine learning to do search, using machine learning to recommend hotel, flight, everything there. I got to see how machine learning is used in the travel industry. Now I’m with Appen, I’m CTO at Appen, where we help our customers to get high quality [inaudible 0:03:42] data to view their machine learning applications. I got the chance to really see a lot of different machine learning applications from all kinds of industries. This is the area I always have a passion and I always say the power and I believe this is a pretty powerful tool to transform all kinds of industries and in the whole world in the future.

[0:04:07] DA: Now, why was now the time to write this book? Were you two inspired by something, was it as simple as you had more time in your hands because of the COVID situation, did you have an “aha moment?”

[0:04:20] ASR: Well, certainly not more time on my hands, I’ll speak for myself. I think Wilson will agree but, you know, I think one of the things that we think this book is quite timely is that we’ve been, the benefit of seeing things in the trenches and watching hundreds of different teams build machine learning based products and sort of attempt to launch stuff and we’ve started to notice a lot of patterns over and over and over again, teams struggling with similar problems. And you know, there’s sort of also a sort of moment in the market here where there’s a lot of hype and a lot of enthusiasm and excitement about machine learning and AI-based products but a big gap and very few companies that are actually using this technology and harnessing it at scale, right? They tend to be larger, more well-resourced technology companies. Others struggle to really get their innovation projects off the ground and I think we saw an opportunity to summarize some lessons learned from the trenches to hopefully brought in access to using this technology to a broader scope of people and help people not stumble on some of the same stumbling blocks that we have.

[0:05:33] WP: I fully agree with Alyssa. I think that the reason which makes us really want to write this book is really see that on the one side, we saw a lot of successful stories how people build AI successfully. On the other side, we see a lot of struggling from all kinds of companies. Whether it’s from big companies or small companies, they ask a lot of similar questions, struggling with similar difficulties. We just believe with all the experience we had before and all the experience we learn from our customers, from our partners, we can help to provide some help to those companies who is struggling with AI and machine learning and also hopefully this is going to help to accelerate the AI adoption across all the industries.

[0:06:22] DA: Now, when you two were writing this book, who did you have in mind that you were writing for? Is this for business leaders only, is this for computer programmers, who could have takeaways from the book?

[0:06:33] ASR: When we were writing this book, we really had folks in mind who were not super technical and did not have a lot of familiarity with machine learning and AI. These are people who may have heard about it at work or might be involved or may be interested in being involved but would like to have some practical guidelines and resources to ask better questions and be better partners with a technology team. On the other hand, this book also might help deeply technical folks who are looking to better partner with their business partners and have a language hopefully that’s more common around how these two groups can work together better. People who are business folks, who don’t have any familiarity with machine learning or some of the technical nuances to get them a little bit more sophisticated in how they’re asking questions and learning some of the lingo. On the other hand, how can technical folks have better fluency in some of the business needs?

[0:07:38] WP: I think Alyssa, sorry, Alyssa covered it well, just to add one little bit of color there. My wife is one of the first readers. She’s from a financial background, has no technology background at all but she can really understand the whole book and resonates with the stories. Just to add a little bit of color there.

[0:07:59] DA: Now, are there similar books like this out there and if so, how is yours different? Is it because of your firsthand experience?

[0:08:08] ASR: I have read quite a few books in the AI space. I have not found one that is similar to the approach that we’re taking. We saw sort of a gap in the market that I think we’re trying to fill with this book. There’s lots of books around how to become a machine learning practitioner from a technical perspective. There’s lots of courses online and lots of technical books that are more sort of textbook geared. Then there’s also kind of, I would say, pretty vague business level books that talk about a story about someone who uses machine learning technology or how it’s breaking down what’s going on in a particular company or particular industry around how they use it, there’s some great books out there but I think we tried to fill a gap here in the market around a practical guide for practitioners who are not looking to become technical practitioners on the business side to try to fill a gap there and do it in the lens of storytelling, around – we interview dozens and include I think over a hundred different real-world examples of stories in our book to try to bring it to life in a way that’s easy to access and understand and not too technical, jargon heavy.

[0:09:30] DA: Now, let’s dive into the book itself and you begin the book with profiles of projects that the two of you have worked on and I have to direct this question to Alyssa first. How does someone who is, as you say in the book, fairly useless at coding, how do you find yourself as an AI leader over at IBM?

[0:09:52] ASR: I was in a different part of IBM and I was working on a big fancy data product that is cool and interesting but I was frankly a little bit bored from a career perspective and I was looking ahead and kind of not interested in optimizing bottom lines for big retailers for the rest of my life and I was sitting next to a woman on a plane, who became a mentor for me and she said, “Go find a hard problem and go solve a hard problem.” I looked around and one of the cool hard problems, this is back in I think 2015, maybe earlier, it was the Watson division at IBM and it was just getting off the ground. It didn’t really have any product leaders yet, very early, it was really a research initiative that they were just starting to put sort of a business lens on and frankly, I got lucky. I had some nice people vouch for me but I applied to be an individual contributor from a product management perspective, which is a little bit of a, sort of, lateral step down for me career wise. I was really excited about this technology and I knew it had this promise and I wanted to learn and so I did and I was very fortunate enough to be paired with some absolute experts in the machine learning space. People who had 20, 30 years of research experience who were willing to white board with me and help me get up to speed and I enrolled myself in Coursera Machine Learning 101 and some other ways to get some crash course. But the value that I was able to ride to these partners of mine, who were experts in the technology of machine learning and the research but not in the productization of it and the creating a business out of it and so that was the value that I could bring as a partner because I had the skills that was very different. What I was able to bring is, you know, a lens and my first job was in the computer division team where I was literally tasked with the strategy of, how do we make money off of this computer vision technology? Do we go after facial recognition, do we go after what’s called natural scene imagery so you know, things you take with an iPhone, do we go after satellite imagery, who were the different people who were trying to harness this to improve their businesses or create outcomes for their customers? That was a new thing and a lens that hadn’t really been put on this technology before at IBM so I got lucky to be on the ground level and help shape that and so through a lot of trial and error, we created quite a bit of revenue for IBM.

[0:12:31] DA: Now, could we stop and just define AI for everybody and maybe some of the real-world ways that’s being used today that maybe people don’t know about?

[0:12:43] WP: A lot of people when they think about AI, they really think like a machine which has human capability, right? We go to Westworld, all those sci-fi movies you see all this AI that people see in their imagination. That’s actually not AI, that’s called AGI. AGI stands for Artificial General Intelligence. They are basically an that intelligent agent who can understand or learn any intellectual task that a human can do. AI today is still far away from AGI so most real-life AI is just an application that has been taught or have learned how to carry out a specific task. It can recommend what you should buy on Amazon, recognize your voice on iPhone or even drive a car for you but unlike a human, it can only learn or be taught how to [inaudible 0:13:38] task instead of a generic intelligence.

[0:13:43] DA: How does it work? How is it generally used? Can it be used in many different industries and businesses? I think a lot of companies are saying, “I don’t even know anything about AI, I don’t think we’re sophisticated enough to have or invest in AI.” What would you say to them?

[0:14:05] ASR: One of the things that I really like a lot from Cassie Kozyrkov, I might not be saying her name correctly but she is the head of Decision Intelligence at Google and she writes, really eloquently, a lot of blogs around machine learning based technology and one of the things that she says a lot is you know, “If you don’t need AI, please don’t use it.” She encourages people not to use machine learning based technology or AI unless they absolutely have to. Because it’s hard and it’s difficult and if you can find a way to solve a problem without it, all the better because often, there’s sort of less sophisticated ways of solving a problem, you know, 80% well. Machine learning based technology is well-suited for a class of problems where all those other statistical or mathematical ways are not able to achieve the outcome that you’re looking for but it’s best suited for problems that are very data heavy and where you have access to large volumes of data that is organized in a way that is easy for machine learning algorithms to consume. As Wilson was saying earlier in his experience at the travel company, you know, a travel company’s great because you can have lots of piles of data in all different sorts of places and often, it’s difficult to extract insights out of them.

[0:15:39] WP: I fully agree with Alyssa’s comments there. AI can be very powerful and it can help things in many ways. It can increase revenue by providing a personalized customer experience or reduce cost by let’s say, automating some customer service tasks. It can help things to be more efficient by analyzing all those alternate data, all their needs. It can be used in sales, marketing, customer service, product, operations. Almost every layer in the business function. AI is powerful and we do encourage these leaders to think hard on how they can leverage AI as their tools to help their business but however, as we discussed earlier, AI is only good at specific tasks so you have to identify the [inaudible 0:16:31] problem in the case you want to use AI. You cannot just expect AI to do everything for you, you just throw a very broad concept to AI that AI will give you the magic. Let’s say ask AI to cure cancer, that’s probably just too hard. You need to have a manually defined specific problem for AI and then a similar team to help you solve that problem. I think that’s the key, key thing for those leaders to understand. AI can be powerful but you have to use it right. Otherwise, it might give you some damages [inaudible 0:17:11].

[0:17:14] ASR: Yeah, absolutely. I think one of the lenses on that narrow and specific problem that Wilson was just referencing is a narrow and specific problem that has a lot of data that you have access to, that has an outsized positive outcome for solving that problem, right? For example, if you are looking to prioritize which customer service requests get escalated fastest, right? You have – hopefully if you're running a customer call center, you have a lot of data, right? A lot of requests and you have an outsized positive impact. If you can answer the most important questions first, that are the most painful or the most urgent for people, you’re going to have a really positive impact on your customer satisfaction scores or you're going to generate more revenue or you can tie it directly to a business outcome that’s really tangible.

[0:18:11] DA: What would you say the goal of the book is for readers? What would you hope they’ll walk away and learn from reading the book?

[0:18:20] ASR: I think the goal is just, you know, deepen their understanding of a topic that they probably don’t have a ton of familiarity with. Come away from reading the book with confidence and excitement and you know, hopefully a little bit of an inspiration about how to use this technology in their own professional lives.

[0:18:39] WP: Yeah, I fully agree with Alyssa [inaudible 0:18:43] and also good on top of that, with this book, we do have hope that readers, they start to know a little bit more about AI and how to do real, AI applications in the real world. Meanwhile, another perspective from this book is also, how can people build AI ethically, responsibly? That is also very important. We want to make sure that all of these AI practitioners not only build AI for business purposes or business benefit but also use AI in a responsible way and bring the good to the society.

[0:19:22] ASR: Yeah, and just to add onto that, I know Wilson and I are both incredibly passionate about making sure that when people are using this technology they’re harnessing it in a way that is fair and ethical and responsible because I think what we’ve both seen firsthand is how easy it is, even accidentally or unintentionally, to encode unintended or harmful bias into a machine learning base system and how critical and important it is to not do that. One of the things that our book addresses is very easy simple ways to avoid doing that accidentally.

[0:20:04] DA: Yeah, you know you walked me into my next question, which was, you do talk about and use the term “responsible AI” in the book and so that led me to want to ask you, what is irresponsible AI and who actually defines that line that you hope wouldn’t be crossed?

[0:20:22] ASR: That is a very big and complicated question that I couldn’t possibly provide a good answer for but I will give the world according to Alyssa here. It’s easy to be irresponsible with AI and I will use my own example of that, that I certainly included in the book, which is that I didn’t know better. I was a little naïve getting into this and you know, we were very close to launching a machine learning computer vision API in IBM that included really negative harmful biases, that described pictures in ways that were unsavory. You know, there’s some public examples of you know, you do a Google image search for the word nurse, right? There’s a lot of female nurses that come back, right? The example that I had was that one of the things that we were labelling people on a wheelchair with was the word loser, which was just horrible. It was not intended to be that way, right? There was no one anywhere at IBM who at all set out to encode that type of a harmful bias but we did unintentionally. We got lucky that we didn’t ever launched it into production and we were able to correct for that and I don’t think there is any specific governing body yet that the laws, at least in the United States, have really not sort of caught up to the sophistication of the technology but I think it is incumbent on anyone who is using this to really think through how – what are some of the ramifications for how this decision getting made and a label being applied or an outcome chosen or a path forwarded down by a machine learning based decisioning engine could have unintended consequences and to put guardrails on how it can be used and who can use it and how scalable it is or isn’t.

[0:22:20] WP: Yeah, you probably can easily tell this is an area that both Alyssa and myself are super passionate about. We shared a lot of stories in the book of how AI put in or introduce some bias, ways the employees really – there is an application review on an AI model basically, you have to check all the resumes and say if the person is qualified for the job or not and later that AI application turn out to be, have bias. Really, they kind of gave the male a much higher success pass rate. Like for females, the passing rate is much lower but there was no intention on even the gender factor in the machine learning model but the output is biased. You can say AI can easily introduce some bias, which creates some harm to this world so we want to make sure that every reader, every AI practitioner, they can really keep this in mind but meanwhile, as Alyssa mentioned, it is also very hard. This is still very active research for both the academic world and also industry world. There is a lot of research around how to detect AI bias, how to avoid AI bias and there’s many companies like Google, like [inaudible 0:23:44] have also started to publish some standard on how to define AI bias, how to measure and give the tools to support those. We are very happy to see all of this effort put in there to help us fight against the AI bias but just keep in mind, when everybody build the AI, this is a super important topic.

[0:24:06] DA: Now, when you take the first step to bring AI into your business, what does it take and how long does it take and what kind of resources does it take to bring AI from concept to real world use?

[0:24:22] ASR: We certainly address this at length in our book from a number of different perspectives to summarize what is a lot more detailed. You know, I think it starts with defining the problem that you want to solve and truly understanding the outcomes that you want to achieve and whether or not AI is an appropriate technology to apply to that problem. Often that is sort of an overlooked step but a really important one and I think once you’ve defined the problem, the next piece of the journey is really looking at the data that you have available to train a machine learning model to solve that problem, to teach it how to solve it. If you’ve well-defined a problem and you have a lot of data, then you can start crafting a team and sometimes you do the team first, or these aren’t necessarily always in order, but you need a diverse team. Not only a diverse team ethnically but you also need a diverse team in terms of skill sets. You need people who are experts in the technical model creation but you also need data engineering and dev ops and QA and you need application developers typically and UI interaction design and product managers. You need a large variety of skill sets often to actually achieve the solution to the business problem that you set out with and often the machine learning model is a portion of the problem and a really critical, important one but not the only one because if the model only works on someone’s laptop and it is not actually deployed into a production environment, it’s really not going to solve the problem and so thinking through kind of the whole set of resources necessary to solve the problem and how are you going to govern and maintain this model once it is in production because I think that is another sort of misnomer about machine learning based products is that, you know, you train the model and then it’s done. That is never the case, or only very, very, very rarely. You are constantly needing to update and maintain and evolve and monitor the model’s performance towards the problem that you are solving. From a timeline perspective, sometimes in a mature organization that has a lot of tools and resources already in place, you can do this entire process in a matter of hours or days. Other times, this is months or years, if you don’t have the skillset or the technology support in place to do this. Huge variants based on kind of where your organization is in their maturity curve.

[0:27:08] DA: Now, have you seen more success at companies who try to build something like this internally or is this something that should probably be outsourced?

[0:27:17] ASR: I’m not aware of a company that takes a totally black and white approach either way. I think in reality almost everyone sort of does a mix of both for different components of this journey. You know, having some resources internally and other parts of the resources that you need externally. For example, Appen, where Wilson still works, you know, it is a data annotation company and it is part of the AI journey but only one part of it, right? There is also infrastructure and hosting stuff. There is also ability to buy off the shelf models as an API that the big players like IBM and Google and Microsoft and Amazon, they host directly. It really depends on the need and the problem that you are trying to solve and whether or not it is strategically important for your company to build or buy different aspects of it.

[0:28:12] WP: Yeah, I want to add a little bit more on top of that. There’s two more things to consider besides all of those important factors Alyssa just mentioned. If these areas are core to your business, or just a supporting function to your business, let’s say if you are a search company like Google or maybe Yahoo, you know, the search ranking would be a core area to your business. It might not be a good idea to really outsourcing them to other company, right? Yahoo made their choice outsourcing like their search engine to Microsoft many years ago and we saw the consequences here but meanwhile, let’s say if you are an e-commerce company and you want to use machine learning to help your own customer service. Customer service probably is not really the core of your business, so that might be the area good to be outsourcing or maybe buy some solutions from external. That is one important factor to consider, is this area your core business or maybe a supporting function? Then you might take a different approach there. Second thing to consider is really the internal AI maturity level or capacity. AI is powerful but AI is also hard. Do you have the skillset? Do you have the experience? Do you have a team to be able to deliver this? If not, maybe getting some outsourcing company or buy some solution from external to start with and along the journey you can learn to build your own capability, and then later on, you can do it by yourself.

[0:29:47] DA: Now, there is a quiz at the end of the book, can you talk about what that quiz is for and maybe if you have other resources available to readers?

[0:29:57] WP: There is an AI readiness quiz at the end of the book and that will help you to really assess your organization's maturity of AI. We have the quiz in the book, you can also go to the Appen website, where you can answer a series of questions and we are happy to give you the assessment of your AI maturity level and then help you to define a strategy for your company. When we talk about AI maturity, there’s potentially five levels. Level one is really the planning level so the organizations start to have some interest to use AI. They may start to help people if they want to, to support different AI technologies and start to identify the first AI use case of the organization and then discuss around how to define the success criteria for the use case. This is basically the pre-pilot stage of the company. Level two is permutation. This organization start to try AI experiment in kind of a data science context. They start to run some QSA through the technology [inaudible 0:31:10] and the first pilot normally is around how to increase their revenue. It could be a product recommendation or maybe price optimization et cetera. At this stage, the company might set up some AI labs or start to recruit some AI experts there, so that’s level two. Level three is around stabilization. Company, they finish a pilot successfully, now have several AI use cases in production. Those use cases are creating value for the business and the team also has the end-to-end experience of developing AI from pilot all the way to production. They probably have [inaudible 0:31:54] excellence with all of the work your AI community so that is level three. Level four is around [inaudible 0:31:02]. After you have a few successful AI use cases, now you want to get AI used across the organization and most of the time at this stage there is a C-level ownership of AI across the company. Also the organization in this stage is really kind of matrix shaped, so they have AI specialist and meanwhile those AI experts also work in different domains and also the company start to look to AI ethics, governance and the risk, et cetera. The last level is really the transformation level. In this level, AI is just part of the business DNA and people are starting to shift from AI project to AI product. AI capability is fully embedded in all of the different product lines or different operational department. They have a profound impact to people, culture, society at large. Those are in summary the five AI maturity levels and if you really want to understand how your organization is doing, how ready your organization is doing to take some AI initiatives, you can come to the website of Appen, www.appen.com and find the AI readiness test that will tell you how – what is the situation of your organization.

[0:33:29] DA: Well Alyssa and Wilson, you know we just touched on the surface of the book here but I just want to say writing a book, which is going to help so many businesses just open up to the world of AI is no small feat, so congratulations on being published.

[0:33:43] ASR: Thank you for having us.

[0:33:45] WP: Thanks a lot, I really enjoyed the conversation here.

[0:33:48] DA: I have one question left. If readers could take away only one thing from the book, what would each of you want it to be?

[0:33:57] ASR: I think if readers could only take away one thing from the book is that AI is accessible. It is something that they can understand and that they can use and even if it is not something that they want or are ready to use professionally, you know, we are all living in a world surrounded by machine learning based products and I would hope that readers don’t think it’s scary and that they can understand it and they know how to interact with Siri or Google home or any of the many, many machine learning based products that they use every day and understand them a little bit better and to not be intimidated.

[0:34:37] WP: The one thing I would want readers to really take away from the book is AI is difficult, challenging but meanwhile it is also powerful. There’s already a lot for through the experience like people can do this successfully and there’s a lot of tapes to help people to be successful with their AI projects. Don’t be afraid to take the challenge and you will learn along the journey and you will be successful and AI can be a powerful tool for you too and for your business.

[0:35:12] DA: This has been a pleasure and I’m really excited for people to check out this book. Everyone, the book is called Real World AI, and you could find it on Amazon. Alyssa, Wilson, besides checking out the book, where can people connect with you?

[0:35:25] ASR: Connect with me on LinkedIn or alyssasimpsonrochwerger.com, feel free to reach out.

[0:35:32] WP: Yeah, just contact me from LinkedIn, you can find my profile there.

[0:35:37] DA: Well, thank you for coming on the show, best of luck with the book and we’ll hope to speak with you again on book number two.

[0:35:44] ASR: Thank you.

[0:35:45] WP: Thank you, it’s really great to be here.

[0:35:48] DA: Thanks for joining us for this episode of Author Hour. You can get Alyssa Simpson Rochwerger and Wilson Pang’s new book, Real World AI, on Amazon. Also, you can also find a transcript of this episode and all of our other episodes on our website at authorhour.co. For more Author Hour, subscribe to this podcast on your favorite subscription service. Thank you for joining us, we’ll see you next time. Same place, different author.