Philip Resnick explains how humans shape AI at every stage: choosing training data, defining goals, and embedding knowledge that can introduce bias and errors. Real examples include medical-record analysis, social-media sorting, and site search, where human experts review outputs and feed corrections back into the system to improve results.
Chapters
Hi everyone, welcome to today's Biznology webinar. We have a great webinar today because I'm joined by Philip Resnick, who's going to be giving us insight into the human element in AI. For those who are new to our webinars, let me give you a bit of background. Our Biznology webinars are half an hour, and our platform allows you to chat with other attendees through the chat tab. You can also tweet about the webinar using our hashtag #BizNoWebinar. Any questions for our speaker can go in the Q&A tab, and we'll ask them at the end. We also can't get started without giving a shout out to our sponsors. Garrus Corp is a full-service digital strategy firm that reaches deeper into the conversation than any other agency anywhere. Thematically, machine intelligence, human insight. Conversion, AI-powered text analytics and insight designed for the social age. And SoloSegment. Revenue is trapped in your site search. SoloSegment Site Search Inspector can help you set that revenue free. So let me tell you a little bit about myself. I'm Mike Moran, founder of Biznology and a senior strategist at Converzion and SoloSegment. I'm the co-author of Search Engine Marketing Inc. and Outside-In Marketing and the sole author of Do It Wrong Quickly. I'm a veteran of IBM, managing groups on ibm.com for 8 years and retiring from IBM in 2008 as a distinguished engineer. That's enough about me. Today we're talking with Philip Resnick. Philip, thanks so much for being here today.
Thank you, it's a pleasure to be here.
So I know, Philip, that you have a great deal of expertise on artificial intelligence. You are an expert in natural language processing and AI with a track record in academia and industry. You're a professor at the University of Maryland in the Department of Linguistics and the Institute for Advanced Computer Studies. You're a technical co-founder of CodeWrite, a company doing natural language processing on healthcare records that was acquired by 3M in 2012. And if you— as if you don't have enough time for anything else, you're an advisor to Conversiyon, FiscalNote, and SoloSegment. And your current startup thematically came out of stealth mode starting in June. That's quite a resume. Anything else you want the audience to know, Philip? How'd you get here?
I think that pretty much covers it. I have a foot in academia and a foot in sort of the industry and the entrepreneurial world, and I like to— that's sort of getting to live the best of both worlds.
Well, it's from that resume, I think you have about 5 feet, so I think that's pretty impressive. But so, hey, Phil, we've known each other a long time, so none of that's a surprise to me, but it was great to introduce you to the audience. So let's get right into the questions. There are a lot of different ways that people define AI. What's your definition?
Yes, there are definitely a lot of ways. The definition that I'm fond of hews pretty closely to what Alan Turing started out with, you know, decades ago at the dawn of computer science. I think of AI as the field in which the goal is to enable computers to do things that we would consider intelligent if people did them.
And that makes sense. One of the problems I always have with that definition is I don't think— I think the hard part of the definition is not defining what artificial is. I think the hard part of the definition is defining what intelligence is.
Absolutely, yeah. And in fact, Turing had an interesting comment somewhere in this discussion where he observed that The assumption that the people in a conversation are both intelligent, whatever intelligent means, is really nothing but what he calls a polite convention.
What does that mean?
It means you have no way of knowing whether I'm really intelligent, but you're polite enough not to tell me so.
All right, well, I'm polite enough not to tell you so, but I've known you long enough that I know you're really intelligent, so I'm going to ask you a tough question first.
Okay.
We keep hearing about deep learning, which from everything I can tell is much better than the shallow kind. And I guess I want to know from you, what is deep learning? And with all the hype we hear about it, does it live up to it?
So yeah, the what is is a very long conversation, but the short version is there is indeed a ton of hype. The fact that it's named deep learning helps because it's deep. That said, there really is a there there. Deep learning is basically a renaissance of approaches to machine learning that date back decades, but are now souped up and practical because of the availability of a lot of computing power, and even more important, the availability of tons of data. Probably the main thing that you need to know about deep learning— a couple of things. Number one, It, unlike a lot of previous approaches to AI, is very, very heavily data-driven, as opposed to needing to build knowledge into the system by hand, so to speak, which was a property of older approaches to AI. And the other thing that I think is quite interesting about it is its ability to automatically learn a lot of useful features or representations of the data automatically. That said, it ain't magic. There is a lot of hype around it, and like any other thing, you actually need to look at the properties of the solution with reference to the problems that you're trying to solve, as opposed to simply, you know, here's a label, everybody says it's great, now let's go for it.
And I get it ain't magic. I mean, to me, magic's a word like panacea. We use it to always say something isn't that. So, I don't even know why we have—
There have been real advances using deep learning for particular problems. So, you know, I was initially very skeptical of the hype, but I think that for a particular class of problems, deep learning has proven itself to be a real advance.
So, we promised that this was going to be about the human element in AI. I guess what I wanted to ask is, so how does human knowledge get into an AI system?
Right. One important thing, and this connects to the deep learning part of the conversation, is that even in systems that are fundamentally driven by learning from data, there is always a human element in that process because the development of the system itself, involves human choices and human knowledge getting built into the system. Probably the simplest example of that is the choice of what data you're going to train a system on. So if you think about face recognition or you think about speech recognition, an interesting illustration of this is the fact that face recognition and sort of the facial image technology does a lot of the time less well for underrepresented groups because it's trained on the well-represented groups. Speech recognition— try speaking to your Alexa if you have an unusual accent. Accent, right? So one way knowledge gets into a system is by the choice of the data that you're going to train a system on. In addition, we build in— whether it's in the form of statistical priors or the features we choose to look at or the structure of our models— we build knowledge into systems, even systems that are, that are learning systems, really from the get-go. And those choices have an enormous impact on the way the system performs.
Yeah, that makes a ton of sense. I mean, I get your example of what kind of training data we feed it. I think it's also true that the human element is involved in just what problems we give it. I mean, for example, we haven't given it the problem of elbow recognition. We haven't thought that that's really an important problem, right? So a lot of it is just the problems we select, but you're right, it's also the data that we select as well.
Mm-hmm.
Those get into questions of bias and other types of things that you're seeing in the news all the time, and all those things are about the human element in AI.
Absolutely. In fact, one of the other things that's worth mentioning there is it's not just about the data you train on, but your definition of success. What all of these learning systems are doing is essentially searching among lots of possible ways of solving the problem. An enormous space and searching for something that is going to meet a goal that is defined mathematically in terms of what you give the system, the objective function or the loss function, it's called. The choice of what that objective is, what are you trying to optimize? Are you trying to simply make it more likely that somebody will have an ad up in front of them if you're serving ads, or are you trying to make it more likely that somebody will click through it, or are you trying to Measure the extent to which it actually leads to some outcome in your system. These are 3 different ways of measuring how well the system did, and the system is going to optimize whatever you tell it to optimize. That's an element where human choice goes into the picture before the actual human use of the system as well.
Yeah, I think that makes a ton of sense. Those are all examples I can relate to because almost all of my background in AI is in marketing, and so it makes sense to me. But the one thing, the thing that—
Uh-oh, Mike, you've frozen on me. I don't know if that's true for other people in the webinar. Uh-oh, we lost you.
Okay.
We can, we can hear you, Mike. I just can't see you. Yeah, there you go.
Okay, okay, that's cool because I couldn't hear anybody. Well, so we need a little AI in our webinar platform because it should know not to disconnect the person talking. But anyway, it did give me a nice voice message that said your video stream has been disconnected, as though that was a really helpful thing for them. So thanks everybody. But so sorry about that. But I guess what I would occurred to me when you were talking about that is that a lot of people, when they think about the human element of AI, they're thinking about programmers or data scientists, but a lot of the human element in AI is actually the expert in the problem. So in the problem you were talking about, you know, a marketing expert who would actually understand what the business value is of solving a problem rather than just a data scientist or a programmer or somebody who's doing the technology part. And so one of the— I guess one question I have for you is, are there particular examples you can think of where the human element is particularly important, kind of some kinds of applications or problems that you've worked on where the human element was really necessary in order to solve the problem?
Yeah, absolutely. So what you just said is absolutely right. When we talk about the human element, you can think of what happens, you know, in the process of creating creating a system, but it's increasingly being recognized that AI is part of a larger ecosystem. So when I talk about the human element in AI, what I'm talking about is the combination of using the technology in tandem with the human in order to do something better than either of them can do by themselves. That's really when we talk about the human. And by the way, there's a Harvard Business Review article from the end of the summer. I think it was the July-August issue that talks about the human element in AI. For folks who are watching this, they might find that interesting. There's been a lot of attention to this recently. So there are a number of examples that I think are good illustrations of this. You mentioned CodeWrite, for example, a company that I helped get started up quite a number of years ago. And that's a place where you had natural language processing and AI looking at medical records. But in order to do a good job of the task, which, for example, might include things like diagnosis or procedure codes, you don't simply want to take what the system has done and then present it to some downstream system without a human in the loop. There are problems where you need a human in the loop to review the system output and ideally to provide feedback in that review process to make the system better. That is the real key. It is not just a pipeline. There is a backward arrow that goes back from what the people did with their human brains to help the system make itself better. Another example of that, you and I are both familiar with Convergion's technology, where again, if you are trying to do things like identify some particular subset of things that are coming out in social media, you do not just want to say, train up a classifier, some kind of learning engine on the basis of labeling data. That will get you up to a certain level of performance, but the definition of what the person is looking for is there in the person's brain. And so it's useful in that context to have technology that lets the person look at what was produced and say, this was good, this was bad, and provide feedback to the system in order to improve what it's doing. The system is great at doing things fast and with large quantities of data, but it takes the human brain in the loop To understand what the problem is they're trying to solve. And the human element and AI together is really about putting those 2 things together and enabling the human brain to provide feedback back to the system. That loop is really what I think of when I use that phrase.
That makes sense. And I mean, another aspect of that that I want to ask about is that people are always fixated on the fact that AI makes mistakes that people never would. How can adding the human element to the process mitigate against that? Yeah.
There is a term that I use. I like to use the term howlers. Howlers, my definition of a howler is not just a mistake, it is a mistake that undermines confidence in the system. Everything makes mistakes. Going back to the medical example, you look at a medical record, it says, type 2 diabetes, but it was really type 1, and we all nod and say, okay, we need to improve the system, say. If you take a male patient and read the medical record and then assign ovarian cancer, that's a howler. That's one where the system can be 99% accurate, but if you've gotten it wrong, it undermines confidence. Again, this notion of an ecosystem that involves the technology, And the human in the loop, right, is not just about the performance of the system, it's about the confidence in the system for performing that overall task. I don't know if that gets to the question that you were asking.
It does. And I think of a howler as being an answer that's so bad that even if you don't know the right answer, you know that's the wrong one.
Yeah, I think that's right. And so there are various things you can do to try to get the human in the loop or the human element in the system to avoid those kinds of howlers. One, of course, is to have a human review process, right?
Yeah.
Or to design the ecosystem in such a way that it's not the outcome of a technology, it's the outcome of a technology plus human partnership, because then ultimately the responsibility lies with the person. Uh-oh, dropped in and out again. Fingers crossed.
I think I'm still here.
Okay, no, I, I can't see your face anymore.
Oh well, it's not that great to look at anyway. I have a, I have a face made for radio, so I think that that's fine. So, well, one of the things that I think adding the human element can do though is it can help us solve problems that AI really can't solve by itself, and, and it actually occurs to me that your startup thematically is a good example of that. Why don't you tell people about how that works and why the human element is actually critical to solving the problem, even though AI is doing a lot of the heavy lifting.
Yeah, thank you. So the tagline, machine intelligence, human insight, is a very short way of capturing it. A slightly longer version of that is that the premise that we have been following is that the role of AI is not to replace human intelligence, it is to accelerate it. So the core problem that a lot of people need to solve with large quantities of language data is, at a very high level, the idea of making sense of it. And what does that mean? That means identifying categories. You have all of this unstructured stuff. How do you actually boil that down into a finite, manageable set of themes or categories in order to understand what's going on, whether that's social media or open-ended survey responses or any other kind of— or the contents of a website. How do you break that down into a set of categories? Doing that fully automatically will get you a part of the way, but we talked about this trust issue, right? If you want to trust the outcome, then often what comes out of the automated systems doesn't get you far enough. And so there is an approach that we have been taking that puts the human in the driver's seat. So, it lets the person drill down, understand why the system did what it did, and which gets to a broader issue of explanation in artificial intelligence, by the way. Explaining what systems do is a really big issue, and it enables the human being to provide feedback. This is good, this is bad, this is signal, this is noise, as feedback to the system, and then recalculate what it's doing to get something that is going to be a better model of what's going on, that's going to be more useful. That's a classic example of that human-in-the-loop backward error that we're talking about, this time specifically designed for the problem of looking at unstructured text and getting useful themes and categories out of it.
Very cool. Yeah, and I mean, you know, I've been working in this area for 30 years, and I know that that's been a problem that's gone back and forth between fully automated solutions that create categories that people think are dumb and solutions where people define all the categories, and then the computer has a really hard time bucketing things correctly. This to me is a great middle ground there. I think it's really cool technology. You're also an advisor for SoloSegment, and I guess they focused a lot on artificial intelligence, especially around site search. I guess, can you tell me a little bit about the process of improving site search? What kinds of things can be automated and what kinds of things can't?
Yeah. I think the real key is to recognize that Improvement in AI comes from data, but data in tandem again with a human element. And so, you know, if you're thinking in terms of a website, any website, the human recognizes what the business problem is. Often that is getting to an outcome of some kind. And so, you know, step 1 is to be able to instrument things, to be able to recognize or identify when an outcome has been achieved, because that's really where the learning comes from, right? The second piece of this— uh-oh, in and out. I'll pause for a second there. Are you back, Mike?
Yes, I am now. Sorry about that. We seem to be having a lot of problems here. So I'm sure you were never more eloquent.
No, I decided to wait, with apologies to the folks that are watching. So basically, I said the first element is actually getting data on the site, so the ability to gather data comprehensively. And in a fine-grained way. But then the other element of this is to be able to recognize patterns in that data that are predictive of the things that you care about. And so there's an element of predictive modeling that is driven by the data. There is an element of insight into the business problem where you have these various ways of getting knowledge into the system, like what features of things are going to be important. To pay attention to. But just as crucially, there is this element of being able to put the learning together with the human insight in a way that's going to be more effective than either of those things. And I think probably the example that I gave earlier where you're saying, well, what ads do we serve to people, the different layers, is a good instance of that. Every step along that path is valuable data, but ultimately, it's the recognition of the desired outcome and recognizing how to get people along the path to that desired outcome that's going to make a difference.
Terrific, terrific. And so this has been a conversation where I've specifically tried to give you— to get you to give very specific examples, and I know some of it has been, you know, probably a little technically dense for some people. And I guess the thing I'd wonder is, what are the things that a business person would really care about that's coming out of all of these lessons, all these examples? Could you sum this up in a way that somebody who doesn't have a technical background, doesn't really understand AI that well, but they really know it's important and they care about getting its value for their business, What are the lessons that they should be drawing out of these examples you're giving us?
So we touched on one at the very beginning, which is that there's a lot of hype around phrases like AI and machine learning and deep learning. I think one thing that I would say is don't be afraid to go a level past the hype, go a level past the buzzwords. Look, here's an analogy. If you were buying a car, Right? Most of us who were like going and buying a car would not simply look at the glossy brochure on the front of it and the little marketing slogans that people were putting. We, you know, I don't— I'm not a person who can repair my own car, but I'm able to go, you know, a level deeper to try and understand, well, if you're paying more for this particular thing, what is that thing? What is the outcome of it? and how does it relate to the way I live my life? There's an analogy there to what I think people should be doing when they're looking at solving business problems. What you need to do is bridge between the problem that you're trying to solve and the technology that purports to be solving it. Don't be asking, does it do deep learning or does it do AI? Be asking, what are the the ins and the outs. I worked with somebody once who liked the phrase, the gazintas and the gazautas, right, of this particular technology. And how do those things bear on the problem that I am trying to solve? That's probably the strongest piece of advice that I can give. Don't be afraid because you have people throwing around terms or you get an academic like me talking and delving deeply, going straight to technical terminology or something. Don't let that put you off from looking at what the technology is actually doing in relation to your business problem. It's that bridge that I think is the most important thing. And I think based on a lot of experience dealing with salespeople and marketing people and so forth, right, there's a temptation for people to say, ah, no, no, no, I can't understand that stuff. You know, get one of the tech guys here, whatever it is. Get the tech guys to explain it to you, and if they can't explain it to you, get somebody who can. Because it's, as we said earlier, it's not magic, right? Go a level deeper. You don't need to drill all the way down into the equations of how it's done. There is a middle ground of relating the business problem to the ins and the outs of what the technology is providing you. That's a very high-level piece of advice, but I think it's very applicable for AI today.
I think that's great advice. And here's another question, just maybe a little similar to it, is what do you think the biggest misconceptions are that marketers and business people have about AI? Oh boy.
Well, I think, you know, it's interesting. I'm going to turn that around on you a little bit. Some of it is one of the biggest misconceptions that people have about human performance. One of the things that people often don't recognize is that they're trying to solve a task that is defined by what people do, But they actually don't understand how well or poorly people do it. If you're getting AI in the loop and saying, I want you to make such and such a prediction, people often will say, okay, so the goal is 100% accuracy. Well, if this is a problem where people can't agree with each other on what the right outcome is more than 80% of the time, then it's silly to be trying to task your technology with getting it right 100% of the time. The fact of the matter is that if 80% was good, if people were doing it, then 80% is probably good if you can do it faster, cheaper, and at a larger scale.
Yes.
And so probably that misconception that the goal is 100% of whatever it is you're doing, right, there is something that recognizing that I think would actually change the way that people approach things. Yes, 100% is great, but most of the time it is not actually the goal. The goal is to do a good job solving the problem that you're solving, and often a good job is defined in terms of people's capabilities and the ability to scale that up.
Well, I mean, that's a great, great thing to remind people because, I mean, you and I have worked on projects where we actually know that the human has a very difficult time actually picking the right answer, even tagging things properly. I mean, we've worked on projects together for SoloSegment and Thematically where there's such a large number of subjects for people to tag that there are studies out there that show that if you give the same person the task of tagging something into a large taxonomy, and then you give them that same task the next day, they will only agree with themselves 65% of the time. And so this is obviously a problem humans aren't very good at, and so to think that machines are suddenly going to be much better than that than humans is something that sets a lot of false expectations. So—
That's right. Always the goal is not to focus on the numbers, it's to focus on the problem. It's to focus on what's going to be valuable and useful. And good performance metrics are always something to pay attention to, but they're not— it's not sort of like a religious issue, you know, this particular metric for this particular purpose. You have to look at it in the broader context.
That makes sense. And I can't think of a better way to end because You can't, because the only people who can understand and focus on the problem are the humans. The machine doesn't know how to value the outcomes, doesn't know what the importance are of solving the problem, and so that's really the basics of how human insight really belongs inside AI. And I apologize, Philip, that's all the time we have today. I also apologize for the technical problems we had, but I want to thank you for all these great ideas, and I also want to thank our audience for joining us. You can sign up for our event emails at biznology.com, and you can receive a recording of this webinar. You can also stay tuned on that same mailing list for an announcement of our next webinar. If you sign up to get emails at biznology.com, we'll be sure to let you know when that is scheduled. I just want to thank everybody for coming, thank Philip again, and just say have a great day, everybody.
Thanks, take care.
