Kirill Eremenko: This is episode number 257 with AI researcher, Melanie Mitchell.
Kirill Eremenko: Welcome to the SuperDataScience podcast. My name is Kirill Eremenko, Data Science Coach and Lifestyle Entrepreneur. And each week we bring you inspiring people and ideas to help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple.
Kirill Eremenko: This episode is brought to you by my very own book; Confident Data Skills. This is not your average data science book. This is a holistic view of data science with lots of practical applications. The whole five steps of the data science process are covered from asking the question to data preparation, to analysis, to visualization and presentation. Plus, you get career tips, ranging from how to approach interviews, get mentors and master soft skills in the workplace.
Kirill Eremenko: This book contains over 18 case studies of real world applications of data science. It covers off algorithms such as Random Forest, K-nearest neighbors, Naive Bayes, Logistic Regression, K-means clustering, Thompson Sampling and more. However, the best part is yet to come. The best part is that this book has absolutely zero code. So how can a data science book have zero code? Well, easy. We focus on the intuition behind the data science algorithms so you actually understand them, so you feel them through. And the practical applications, you get plenty of case studies, plenty of examples of them being applied.
Kirill Eremenko: And the code is something that you can pick up very easily once you understand how these things work. And the benefit of that is that you don’t have to sit in front of a computer to read this book. You can read this book on a train, on a plane, on a park bench, in your bed before going to sleep. It’s that simple, even though it covers very interesting and sometimes advanced topics at the same time. And check this out, I’m very proud to announce that we’ve dozens of five star reviews on Amazon and Good Reads. This book is even used UCSD, University of California San Diego, to teach one of their data science courses. So if you pick up Confident Data Skills, you’ll be in good company.
Kirill Eremenko: So to sum up, if you’re looking for an exciting and thought provoking book on data science, you can get your copy of Confident Data Skills today on Amazon. It’s a purple book, it’s hard to miss and once you get your copy on Amazon, make sure to head on over to www.confidentdataskills.com where you can redeem some additional bonuses and goodies just for buying the book. Make sure not to forget that step, its absolutely free. It’s included with your purchase of the book, but you do need to let us know that you bought it. So once again, the book is called Confident Data Skills and the website is confidentdataskills.com. Thanks for checking it out and I’m sure you’ll enjoy.
Kirill Eremenko: Welcome back to the SuperDataScience podcasts ladies and gentlemen, super excited to have you back here on the show today. And the guest for today is Melanie Mitchell, who is a professor at Portland State University and author of six and soon to be seven books on the topic of artificial intelligence, and online course creator, and one of the leading researchers in the field of AI. And what you should expect from today’s episode is a very chilled, laid back, relaxed conversation about AI complexity and supporting topics.
Kirill Eremenko: So we’re going to go into a few philosophical areas and what you’ll hear about is complexity, what it is and how it works, and how it can be seen in different areas of life from ant colonies, to the human brain, to the Internet itself. We’ll talk about common sense, meta-cognition, explainable AI, what it is and what the trade-off is with efficiency of artificial intelligence. We’ll talk a bit about DARPA and military applications of artificial intelligence and you’ll also hear Melanie’s ideas and thoughts on the future of AI, which break down into two areas which you’ll find out in this podcast.
Kirill Eremenko: So quite a philosophical conversation coming up, and before we dive straight into it, I’d like to do a shout out to our fan of the week who is Joseph and who said, “This series is truly informative. I have just started to take the first steps in data science and this podcast not only helps to learn the basics, but keeps us informed on the latest trends in this field.” Thank you very much, Joseph. I’m sure you’re going to enjoy this particular episode. And for those of you who haven’t yet left a review, you can head on over to iTunes or to your favorite podcast App, and leave your comments there. I’d love to read them and get to know what you have to say.
Kirill Eremenko: On that note, let’s dive straight into it, and without further ado, I bring to you Melanie Mitchell, a leading researcher in the field of artificial intelligence.
Kirill Eremenko: Welcome back to the SuperDataScience podcast ladies and gentlemen, super excited to have you on the show today because with me, I have Melanie Mitchell calling in from Portland. Melanie, how are you doing today?
Melanie Mitchell: I’m doing great. How are you?
Kirill Eremenko: I’m well, thank you very much. Super pumped to have you on the show. I’m very, actually, just as we were talking before, very excited to talk about all these topics about your books, about your courses, about the work that you do, complexity, artificial intelligence, and all these other areas. Probably to get us started, can you tell our listeners please, who is Melanie Mitchell and what is it that you do?
Melanie Mitchell: Right. So, I do research in artificial intelligence and machine learning, and complex systems. I’m a professor at Portland State University in Oregon and I’m also external faculty at the Santa Fe Institute in New Mexico. And I work on both research and education, and writing. So, I do a lot of writing and I have several books on these various topics.
Kirill Eremenko: To be more precise, Melanie has six books and one more coming out later this year in September, I think you mentioned. Congratulations. It’s so exciting.
Melanie Mitchell: Yes. Thank you. I’m excited about it.
Kirill Eremenko: And in fact, one of your books; Complexity: A Guided Tour. Won the 2010 Phi Beta Kappa Science Book Award and was named by Amazon as one of the 10 best books of 2009. Is that right?
Melanie Mitchell: Yes. One of the 10 best science books.
Kirill Eremenko: Best science. Yes, best science book of 2009. Tell us a bit about that book; Complexity: A Guided Tour. Let’s start with complexity. What is complexity?
Melanie Mitchell: So, complexity is a very broad area that deals with what are called complex systems, which are systems that you can say they’re more than the sum of their parts. So think of the brain, for example, which consists of hundreds of billions of neurons, each doing some relatively simple operations. But together, somehow emerging out of that giant system is what we call intelligence and oceans, and cognition, and all of that. And so the question of complexity and there’s other complex systems like the economy, the immune system, insect colonies, people are looking at what are the commonalities among all of these systems? What can we say about complexity in general across lots of different disciplines?
Melanie Mitchell: So my book is an overview for a general audience about what complex system is, what has been done in the field, what are the big questions and why is it all important?
Kirill Eremenko: Interesting. So, what would you say is like one golden nugget that you can share from your book with us today?
Melanie Mitchell: So one of the things I talk about is the science of networks. This is a very general area in which people look at how networks, that is huge collections of entities that are linked together in some way. You can think of a computer network or the brain with neurons being linked together, or a social network. How are these networks, structured and is there anything in common that makes networks in nature and maybe in technology also work the way that they do? What makes it resilient?
Melanie Mitchell: And it turns out that there’s in the last maybe 30 years, there’s been a lot of discoveries about commonalities and universal laws regarding these networks. And it’s just fascinating that something like the Internet has some properties in common with the brain and it has properties in common with economics. And the question is why? How did these things come about and how are they resilient? How are they vulnerable? Yes, and so on.
Kirill Eremenko: Interesting. So it’s almost like a template for an entity that is applied across different areas we see, whether it’s internet and colonies of human brain, but there’s something in common across. And so, by discovering features in one area, we might be able to see them and apply them or leverage them in other areas of life.
Melanie Mitchell: Exactly. Right. Yes.
Kirill Eremenko: Very interesting. Somebody recommended me this book, but I haven’t read it myself. Just wanted to get your opinion on this, have you read “The Square and the Tower: Networks and Power, from the Freemasons to Facebook”?
Melanie Mitchell: No, I haven’t read that but it sounds fascinating.
Kirill Eremenko: Yes, it would be pretty cool for both of us to read and talk about it. It sounds like you’d be the perfect person to discuss it. But anyway, let’s get back to your book. So you have this Complexity: A Guided Tour, and what’s the book that’s coming out in September. You mentioned that that might be the most relevant one for our audience.
Melanie Mitchell: Yes. That’s called Artificial Intelligence; A Guide for Thinking Humans. And it’s a broad overview of modern day AI, through how do some of the most prominent systems that we all use or we hear about, how do they work. What can they actually do versus what are they claimed to do in the media? How far are we now from human level AI and what even does human level AI mean? So the book is really, it combines both philosophical discussion with actual, getting into the details of how deep learning works and how programs like Alpha Go, which a recent program that beat one of the world Go champions. How does all that work and just how intelligent are these systems really? So it’s really meant to be an accessible exploration of modern day AI and some of the big questions surrounding it.
Kirill Eremenko: Also, now, I know the book’s not out yet, but is there anything you can share with us to give us a teaser or a taster for what to expect inside the book?
Melanie Mitchell: Yes. So one of the things I talk about is the idea of narrow versus general AI. Do you know what we have? We’ve seen a real revolution, you might say an AI over the last 20 years, where systems, including deep neural networks, have become incredibly good at certain tasks like speech recognition, object recognition in images, playing games like Go and chess, and so on. But these are all pretty narrow areas, like Alpha Go is the best chess player in the … I mean, sorry, the best Go player in the world, but it can’t do anything else, it can’t play any other game, even. It even can’t play any slight variation on Go.
Kirill Eremenko: Let alone cooking breakfast or something like that.
Melanie Mitchell: Right. And the question is what would it take to get a system that would be more general like humans are? And I think humans often don’t even know all the things that they actually are good at that computers are actually very bad at. Because things come to us so easily, like for instance, just having general common sense, being able to describe what we see in an image, being able to take something that we learned, like playing checkers and transfer that to some very similar games. How does that all work and why can’t current machines do that? That’s one of the big things I talk about in the book.
Kirill Eremenko: Interesting. So when I was learning about artificial intelligence, what I’ve found about neural networks interesting, is that they’re designed in a way to mimic the human brain. But at the same time, they’re much more, as I understand, they’re much simpler, much more basic than even the neurons that we have in our brains. Would you agree with that and do you have any additional comments on that?
Melanie Mitchell: Yes, I agree with that. They’re inspired by the brain, and in fact, now they’re called neural networks after all, but there’s a lot of important differences. One big difference is that most of the most successful neural networks are what they call feed forward. Meaning that the input goes in one end and it moves through layers of the neural network in one direction up to the output, but there’s no feedback. Whereas, in the brain, especially say in the visual system, there’s 10 times as many feedback connections as there are feed forward connections.
Kirill Eremenko: Interesting.
Melanie Mitchell: When you look out at as some kind of visual scene, not only is the light coming into your eyes and going… being processed up through the layers of your brain, but expectations and knowledge, and emotion, and all of that is also feeding back to affect receipt. And that’s something that’s almost entirely lacking in today’s neural networks. That seems to be incredibly important.
Kirill Eremenko: Oh, well, how about backpropagation?
Melanie Mitchell: Backpropagation is not the same thing because backpropagation is a learning method where you look at the error that a network made on some example that it was given and then change the weights to make the output more correct. But that’s a one step at a time learning method. But what I’m talking about is just not when the network’s learning, but when it’s actually doing something like identifying an image as somebody walking a dog, right?
Melanie Mitchell: See, look out and you see somebody walking a dog. Okay. You recognize the objects in the image and you know something about these concepts and the whole, process of recognition, aside from learning, aside from backpropagation, that involves a lot of feedback in humans of things that we already know, things that we expect to happen, we can make predictions about what’s going to happen next. And that helps us in making sense of what we see. So perception itself in humans in the brain is a very dynamic process, whereas in neural networks, it’s a very static process.
Kirill Eremenko: Interesting. Interesting. So sounds like even though metrics are inspired by the human body, they’re quite away from what the human brain is capable of.
Melanie Mitchell: Yes, that’s right. And I think everyone in the field would quickly acknowledge that’s true and say there’s a lot more to be done in the field to make neural networks more brain-like and there’s obviously a lot of research towards that goal.
Kirill Eremenko: Got you. There’s about a hundred billion neurons in the human brain. How many neurons do we get up to in neural networks these days?
Melanie Mitchell: Well, I guess you have to have some caveats there. So there’s maybe a hundred billion neurons, but there’s also trillions of connections between them. There’s also other cells in the brain besides neurons that maybe have a lot of functionality. And the brain also has a lot of, not only electrical like neurons firing, but also chemical communication. So it’s quite a bit more complicated than any neural network. I don’t know how many neurons are in the largest neural network today, but it’s almost like comparing apples and oranges.
Melanie Mitchell: And people sometimes say like, “Oh, in the exponential growth of hardware, we’re going to be able to match the computational properties of the human brain in 10 years.” But I think that’s missing a lot about the complexity of the brain and how it’s wired up, and how it operates, how its dynamics work. A similar thing happened with I think the human genome. People thought that once the genome was sequenced, we’d understand quite a bit about how living systems work. But it turns out that the complexity wasn’t in the number of genes, just like the complexity in the brain isn’t the number of neurons, but it’s really the inner connections among them. So there you go with a network, example of a network.
Kirill Eremenko: Interesting. I get, another network. Yes. Also, sounds like even if we increase the sizes of neural networks, there’s other considerations that might be necessary in order to achieve general artificial intelligence one day.
Melanie Mitchell: Yes, absolutely. I don’t think there’s any controversy about that, at least in the field. It’s not just a matter of adding more and more neurons and more and more layers, but there’s some other fundamental aspects of how the brain works, how learning works and so on that we’re really missing in today’s neural networks.
Kirill Eremenko: Well, I guess that’s good news, especially for researchers because that’s where we get new inventions coming up all the time. Though by the likes of the general adversarial networks or the recent publications by Geoffrey Hinton, or the full world model, things like that, where people are experimenting with different approaches that are not the standard, just grow your neural network size. And on that note, I wanted to switch to your research a little bit. Tell us a bit about what is it that you do at your research? First of all, how big is your research lab then what do you guys focus on?
Melanie Mitchell: So, I have about six PhD students working with me and a number of masters students, and some undergrads. And what we’re working on is we’re working on right now vision, how is it that a machine might be able to make sense of visual input in an image and not only how to, for instance, recognize all the objects in an image, but also to have a system that could make sense of all the relationships among the objects. For instance, I mentioned the idea of an image of a person walking a dog. Today’s neural networks can do a good job of recognizing objects in the image. They could recognize that there’s a person, there’s a dog, it might be a leash, it might be in a park with some trees and so on.
Melanie Mitchell: But it would be… it’s often hard for a neural network to recognize those were the relationships that we would recognize that yes, the person is actually walking the dog and they’re walking, and they’re going in the same direction and that they’re kind of connected to each other. And in general, this idea of being able to recognize more complex visual concepts is difficult. So my work is on integrating deep neural networks with representations of knowledge. So prior knowledge that a person might have about concepts and being able to recognize these more complex visual concepts in an image or video we’re also looking at. So it’s integrating neural network approaches with more old fashioned AI, more symbolic approaches.
Kirill Eremenko: Okay, interesting. What pops to mind here is that sometimes as humans, we make… like definitely we’re better at recognizing dogs and people in parks, and predicting where they’re going. But sometimes as humans we make mistakes in recognition. For instance, like if you’re looking straight and then with your side vision you see a shadow, sometimes you might think it’s an animal or it’s a threat to you, or something like that. Or there’s lots of these visual illusions where you’re looking in the center of an image and it looks like the image is moving but it’s actually not moving.
Kirill Eremenko: So in that sense, AI might actually be better than us. So do you think that’s a problem in our brain or is that something that we can leverage in research to understand better how the brain works? Why does it make these mistakes?
Melanie Mitchell: Oh, that’s a great question. Yes, so humans definitely make errors. We’re susceptible to visual illusions. We see faces everywhere, even when there are no faces actually there. We [inaudible 00:24:21] what people call cognitive biases.
Kirill Eremenko: May I just add one more thing before… while we’re on this, sorry to interrupt. I just remembered, I noticed this one really peculiar mistake or something that I was [inaudible 00:24:37]. If you try to look at a human’s face while they’re talking upside down, like I don’t recognize people. Like a family video’s playing and I’m lying upside on in the couch, I don’t recognize myself, my brothers, my parents completely. As soon as I go beyond the 90 degree tilt, it’s completely different people. That blows my mind. Is that just were not designed to look at people upside down? Is that why?
Melanie Mitchell: That’s interesting. Yes, I think that there’s something very specific about faces in our brains, we are really attuned to recognizing faces, to looking for faces because we’re such a social species. And so, I think when you’re looking at someone upside down, you’re trying to make sense of their upside down face as a right side up face, and it doesn’t quite make sense. So I think other objects, we don’t have that problem so much. Faces are just this weird thing that we have. We have some specific brain hardware specifically for recognizing faces.
Melanie Mitchell: But what I was going to say about cognitive biases, the mistakes that we make. So some people have proposed that AI systems will be better because they won’t make the same mistakes we do. They won’t have the same biases. And that’s true in one sense, but in the other sense, it’s not totally clear to me that you can have general intelligence without having these biases.
Kirill Eremenko: Interesting.
Melanie Mitchell: Yes. And I know I can’t prove it, but people talk about super intelligence, machines that are smarter than humans in every way and don’t have the same biases, aren’t influenced by emotions the way we are. And therefore, and can read a billion books in an hour and all of that. I have a suspicion that that’s not possible and that we can’t have it both ways. We can’t have general intelligence without some of the biases that we ourselves have.
Melanie Mitchell: So, I think that’s something that’s going to be… people were going to be examining over the next many decades of trying to understand human intelligence and trying to get AI. I think that there’s going to be a trade-off between general intelligence and being able to be unbiased in this sense. So, that’s just in relation.
Kirill Eremenko: Very interesting. So have you noticed any of these or any kind of biases popup in your research so far?
Melanie Mitchell: Oh, that’s a good question. There’s absolutely art in some sense, our systems aren’t smart enough to have the same kind of biases that people do but our systems, they… one of the things that they have is they have expectations. So because they have some prior knowledge, so sometimes if my system sees a person and a dog, they look for a leash, the person holding a leash sometimes hallucinate it. So that’s sometimes a problem.
Kirill Eremenko: Okay. Got you. Interesting. So tell us a bit about the world of research. What is they’d like to do research versus doing applied artificial intelligence in business, in industries. What are some of the commonalities or significant differences you would say?
Melanie Mitchell: Yes, it’s very different I would say. And usually when you’re having applied, doing some kind of applied work, you have a very well formulated problem. You have a data set, you want to cluster it or find certain communities, say, in a set of data like certain people who have very similar tastes or something. And you take some method that you’ve already had experience with like clustering and you apply it to this data. You try and interpret the results.
Melanie Mitchell: So in research it’s more like you have to come up with the question itself and there might not even be any method that exists that addresses your question. And your results might end up being completely wrong. Your hypotheses might be completely wrong and so at the end of the day, but in research, that’s the normal state of affairs, is that you’re wrong. Whereas, I think in applied research, if you could get a bad result, that’s actually a bad thing.
Melanie Mitchell: So, different people, I’ve had students who came from business and just found it, wanted to do something that they felt was more creative. And I’ve also had students that really don’t like the open ended nature of research. They want to do something that has a more obvious immediate impact and that has a right answer. So, I think it’s not as black and white as I’m putting it because there’s obviously a continuing of activities that people do between research and applications. But, I think my constant state of affairs, it’s being wrong and that’s what research is, that’s what you are in research and that’s okay. That’s part of it.
Kirill Eremenko: Yeah. Got you. It already takes one time to be right to make a technological revolution, and it’s okay if you had a hundred times wrong before then. Interesting. And so do you think, do you have application in mind when you’re doing your research or is it research for the sake of advancing science and then we’ll see how we applied when we get there?
Melanie Mitchell: I think there’s both. I mean, I got into the field because I was really interested in what intelligence is, which is a broad question. And so, I wanted to understand intelligence and one way to understand it just by trying to create it. I can imagine applications for some of my work, there’s all kinds of applications for visual understanding. And in fact, some of my students have gone on and worked for companies and applied some of these ideas. But application, building a system that other people can actually use for real problems is a huge undertaking in itself. Even if all the ideas are worked out, just building a production system is a huge job that I haven’t myself done. So I’ve been focusing more on basic research. Yeah.
Kirill Eremenko: Got you. Got you. Okay. All right. And I wanted to talk a little bit about your piece in the New York Times. So far, listeners in November last year and New York Times published a piece by Melanie called Artificial Intelligence Hits the Barrier of Meaning, and the subtitle is; Machine Learning Algorithms Don’t Yet Understand Things The Way Humans Do With Sometimes Disastrous Consequences. Could you give us a quick overview of what this piece is about and what prompted you to write it?
Melanie Mitchell: So this piece is… it was in response to a lot of the media coverage on AI that we’ve seen. We see headlines such as Machines Are Now Better Than Humans At Object Recognition or Machines Have Surpassed Humans At Reading Comprehension, seen that kind of thing. And these different views of machines are better than humans at playing the world’s most difficult game, Go. And so, these are all in some… you could argue that these are true because the machines have surpassed humans on some particular dataset, benchmark data set.
Melanie Mitchell: But it’s not really true in general because the machines, they can do things like translate from one language to another. We’ve seen translation programs say or recognize speech really well, but they don’t understand in the sense of human understanding what… they don’t understand their inputs are their outputs. And the reason why it can have bad consequences is that it turns out that this makes the machines fairly fragile or some people call it brutal, meaning that they do really well as long as they have the kind of data they’ve been trained on. But if the data changes just a little bit, they can completely fail.
Melanie Mitchell: And also, the people have shown that they’re now very vulnerable to hacking. I don’t know if your listeners have seen these what’s called adversarial examples. A hacker could change an input say to a speech recognition program just a little bit in a very targeted way, change the audio signal. And it would not sound any different to a human, but the machine would interpret it as something completely different and possibly something that you might not want the machine to interpret it as.
Kirill Eremenko: Or that example with the stop sign and a few stickers on it can make the machine see it’s like a 60 kilometers per hour speed limit.
Melanie Mitchell: Exactly. It turns out that these systems are very vulnerable in vision and language, in playing games even, face recognition. And so, if we start having broad applications of these systems that have this kind of fragility, and I argue in the piece of the fragility is precisely because they don’t understand in the sense that we understand these concepts, it can have dangerous consequences. And we’ve already seen that in face recognition, for example, where systems can be fooled pretty easily. And now certain organizations are using face recognition as a critical security method.
Kirill Eremenko: Like iPhones right now.
Melanie Mitchell: iPhones and I think some police forces are using face recognition as a way to catch fugitives or spot criminals, whatever. But it’s not very robust because the system doesn’t have the same kind of understanding of the world that we have. And so, that was the point of the piece. It’s a cautionary note on all, the AI Revolution, it’s real. AI has progressed a huge amount, but it’s still quite fragile in a sense because it hasn’t progressed enough in some sense.
Kirill Eremenko: Interesting. On that whole notion of people being able to fool AI, I was listening to a talk by Ben Taylor recently on YouTube and I think it was in that talk that I heard the notion that as long as we have people who are in combating criminals who are trying to create algorithms that are smarter than what hackers and other people with malicious intent are trying to create, we’re always going to do… it’s always like a double sided coin. You have people creating these protective algorithms, but that means the knowledge about how they work and about where they’re going is out there.
Kirill Eremenko: And not to say that those same people are going to go out and be malicious, but it’s the knowledge is out there and it’s potentially accessible. And that means that somebody can always be a step ahead anyway. And so, as long as we protect ourselves from hacks and all these malicious events, the more we do that, the more stronger and sophisticated the hacks and malicious events are going to become anyway.
Melanie Mitchell: Yes, I think that’s probably right. But then there’s a biological analogy in that, all living things are attacked by other living things. There’s biological arms races all over the place. So, we humans have these very complex immune systems that protect us from most of the things that attack us, but not everything, of course. No, they’re not perfect. But the state of AI right now is that it’s ridiculously easy to attack these systems. And even without attacks, they’re quite fragile. Even if they’re not being attacked, they run into some situation that they haven’t been trained on that they have a problem.
Melanie Mitchell: And sometimes we’d like to get AI to a part of general intelligence I think is being more robust to attacks of any kind. And acts are never going to go away, but living intelligence system seemed to be more robust than our current AI systems to being fooled or being attacked in these ways. And so, we’d like to just increase the amount of robustness.
Kirill Eremenko: Interesting. And so, do you think that how can we and on the other hand, can we even make machines understand meaning better?
Melanie Mitchell: Yes. That’s an open question. So one of the big things that people talk about nowadays is common sense, that’s become a buzzword in AI. People say one of the problems with AI is it doesn’t have common sense and common sense can mean many things. But one of the things that people mean is that we humans have vast knowledge about the way the world works and that knowledge is used in our perception of things that occurred in our lives. We know that if you drop something that’s made of glass onto a tile floor, it’s going to shatter. We learn all kinds of things like that, people call intuitive physics or also intuitive psychology, how people are going to react.
Melanie Mitchell: I know that if I drop a piece of glass onto a floor and it shatters, people will be startled. And we learn all that, some of it is innate, probably, some of it is learned when you’re very, very young. And how do you get machines to have this general understanding of the world? And there’s a lot of funding now. DARPA, for instance, which is one of the biggest funders in the U.S. of AI has a big program called Machine Common Sense where their goal is to get machines to have the common sense of an 18 month old baby. And that’s seen as a grand challenge now. And that’s part of the effort to making machines understand the meaning of the situations they encounter.
Kirill Eremenko: DARPA is part of the military, is that right?
Melanie Mitchell: That’s right. The Defense Advanced Research Projects Agency.
Kirill Eremenko: Interesting. So what are your thoughts on government’s investing more and more funds into defense in the space of artificial intelligence? And has that got any dangerous consequences in your mind?
Melanie Mitchell: Yes. I mean, at least in the United States, the Defense Department has always been the biggest funder of AI. And DARPA has been one of the biggest funders in the defense world and in fact, they have set up their grand challenges for AI that have really pushed the field forward. So, it was their grand challenge for autonomous driving that’s really pushed the whole field of self-driving cars. Their grand challenge on speech recognition that really pushed to the advances in speech recognition. So they’ve done a lot of great things for the field. They really pushed it.
Melanie Mitchell: On the other hand, I’m quite worried about military applications of AI, especially autonomous weapons that would presumably make decisions about who to kill or what thing to bomb, and without any human input. That’s something that I think the military would like to have but I think it’s very dangerous for the same reasons that I talked about in my New York Times Op-Ed. That these systems, they don’t have the same understanding that we have, and I think that presents a lot of danger.
Kirill Eremenko: So like what’s an example of that? What’s an example where a system doesn’t have the same understanding as we have? Even though like we’ve programmed it, we’ve created it and we are quite sure that is going to do as we’ve told it to do, as we’ve pre-programmed it. Do you have any examples where that could back and backfire?
Melanie Mitchell: Well, one of the problems is that we… I mean, what you said, we pre-programmed it, but the way that systems, the most successful AI systems work today is that they learned from data. We don’t program them. They learn from huge amounts of data. And we don’t understand how they make their decisions because the system consists of some deep neural network with millions of weights and it doesn’t explain itself. So it can’t explain to us why it made the decision it made.
Melanie Mitchell: Just like the example you gave with the stickers on the stop sign, why did the system think that that was a speed limit sign instead of a stop sign? Well, it can’t really explain why and people are still trying to figure out how these adversarial examples fool these networks, they don’t totally understand it. And so, we have these systems that work… seem to work really well on the data that we test them on, but we don’t understand how they work or and we also can’t predict where they’re going to fail.
Melanie Mitchell: That’s another question in the whole field of AI to have more explainable AI systems that can explain their reasoning, and that’s very difficult. That’s something we had in the old days of AI when you had expert systems and they could explain because they used human programmed rules, but they didn’t work very well. And so, now we have these systems that work much better but they’re much less explainable.
Kirill Eremenko: So this is like a trade-off, right? Like if we want them to be explainable, we’re risking of stifling AI growth. Not as stifling, but, there’s always going to be, due to the nature of competitive markets, there’s always going to be countries or companies that are developing not explainable AI and they’re going to get ahead. Is that about right? Like at the moment, is it a trade-off between explain ability and efficiency?
Melanie Mitchell: Yes, I think that it can be. And there’s also a philosophical question of what does explainability actually mean? Like, so for instance, the European Union now has some laws about-
Kirill Eremenko: GDPR, we all love GDPR.
Melanie Mitchell: Yes, it has some laws. And one of the things in that is the right to an explanation, I think it’s called. Or like a computer system tells me that I can’t have a loan that I applied for. I have a right to an explanation, but what does that even mean? What’s an explanation? You know? So, if I tell you all of the weight values and my neural network, is that an explanation? Well, not really because a human can’t understand anything about it. But it’s not clear what constitutes an explanation. So, think that’s what kind of a philosophical issue that [crosstalk 00:47:43]-
Kirill Eremenko: What would you say expandability is? You are one of the leading researchers in this field. If anybody, you should have the answer.
Melanie Mitchell: I think it explainability as we know, depends on, it’s kind of a social construct. I’m going to explain something to you. I have to have some theory of mind of you. I have to have some model of what your prior knowledge is, what level of explanation you’re looking for… And it’s really a social thing, I think explanation, and that’s something that we don’t really have with machines, is that whole social component. The machines don’t have any intuition about the people that they’re dealing with, or how to explain… explanation is very context dependent, let’s say.
Kirill Eremenko: Interesting. So, basically we need another AI to explain AI to humans. I mean explain that.
Melanie Mitchell: That mean something that people are working on, is what people call metacognition, which is cognition about cognition. So, understanding your own cognition enough to explain it to someone.
Kirill Eremenko: Interesting. How far ahead are we on that front?
Melanie Mitchell: Not very far. And people aren’t always good at this either.
Kirill Eremenko: Yeah. Oh, that’s so true.
Melanie Mitchell: People will give explanations that really have nothing to do with the real reason.
Kirill Eremenko: Yeah. I’ve heard that and I’ve probably done that many times myself.
Melanie Mitchell: You don’t even know that you’re doing that, but people will… It’s been shown many times in psychology experiments that people rationalize a way things that they did after the fact. And don’t even consciously know why they did a thing.
Kirill Eremenko: Melanie, I also had this, a recent revelation and I wanted to run this by you. With humans, I always thought that like our brain is the main source of all of the thoughts and actions and so on. And then like, and then that goes down the body and the rest of the body is just mostly for executing and surviving and keeping, keeping the brain running. And there’s this kind of like a show, kind of like a cartoons being around for a while. It’s called Futurama. Have you, have you seen Futurama?
Melanie Mitchell: Yeah, I’ve seen it.
Kirill Eremenko: Yeah. So, they have this one character, I think it’s, the preserved Richard Nixon from back in the day, but just his head, then they put him on the robot and then he moves around and like, and can think and so on, and kind of like, that’s pretty cool. But what I learned recently is that a lot of our emotional state is actually directly connected. There are nerves that go straight from the core of our brain to or go to the core of our brain straight from our intestines from stomach and other smaller, larger intestine and so on. So, basically your gut flora affects directly how you are feeling and what mood you’re in.
Kirill Eremenko: So, it’s actually a much more complex structure than just the brain itself. And with that in mind, will machines ever like, even if we were able to recreate the brain, there’s so many other aspects to human emotion and cognition understanding, meaning, will machines ever be able to understand this or once we do create them and give them that capacity to see meaning, they will just never be able to relate to the same way that we do to events and objects and things that they see and hear and experience?
Melanie Mitchell: Yeah. I don’t know the answer. I think it’s a good question. There’s a branch of AI, it’s called embodied cognition. Which the hypothesis is that it’s ridiculous to think of this idea of a disembodied brain, which is what most AI systems are, without having a body-
Kirill Eremenko: Thank you for putting into scientific terms, what I tried to just describe.
Melanie Mitchell: Yeah. I mean it’s completely valid. I think there’s a lot to it that we don’t realize. People, we see the brain as being this central processing unit and everything else has kind of peripheral, but it’s really not correct, because biology figures out more and more about the complex systems. That is the body, we’re going to see that there’s so much more to thinking than just neurons firing in the brain. I think you’re absolutely right.
Kirill Eremenko: Interesting. So would you say that the… I’m just curious to get your stance on the whole issue. Some, researchers and scientists say that AI, general AI, as soon as it gets here it will be a massive help to us and save lives and help us invent things and propel the humanity forward. And others say that one’s general AI, gets here, it will completely not understand humans and think that we are a plague on this planet and wipe us out. What, what, what are your thoughts on these two?
Melanie Mitchell: I think we’re very far from understanding what general intelligence is. So, it’s really hard to say what general AI would do or wouldn’t do, or be like. I think we underestimate the complexity of intelligence, our own intelligence, which is why we think that a lot of people think that general AI is imminent. I don’t know the answer cause I don’t really think we understand enough about intelligence to say what would happen. I think there are dangers that we should be aware of.
Melanie Mitchell: But one of the things I quoted in my Op-Ed was, Pedro Domingos, who’s an AI researcher from University of Washington. He had a book where he said, “The real danger… ” I can’t remember the quote exactly, but it’s like, people say that AI is going to get super intelligent and take over the world, but the real problem is it’s actually too stupid and it’s already taken over the world. We trust too much in AI that’s not smart enough. Rather than being faced with the danger of too smart AI.
Kirill Eremenko: Interesting. Very interesting quotes. When you think about it, the technology that we use is already the extension of our lives. Like we look at our mobile phones like 150 times per day.
Melanie Mitchell: Yeah.
Kirill Eremenko: It’s harder to imagine walking outside the house without your mobile phone. It’s ridiculous.
Melanie Mitchell: Yeah.
Kirill Eremenko: Very interesting world we live in. Melanie, on that note, I actually had just one more question for you. From, from the perspective that you have and from what you’re seeing at the like at the forefront of artificial intelligence. Are there any, or is there any recommendation you can give to all listeners who are data scientists, aspiring data scientist or business managers and owners? What’s to look into what to be prepared for in the future of AI in the coming one, two, three, maybe five years at most?
Melanie Mitchell: There’s a couple of things. One is that, the whole connection between AI and cyber security is getting, more and more strong that, that AI, as it gets more capable and more broadly deployed, becomes more vulnerable to attacks. And that’s something that people are just beginning to grapple with. And some of the cyber security people have been talking about this for many years, but I think people in sort of the real world of AI applications are just beginning to grapple with the security implications.
Melanie Mitchell: Another thing is that, that I think there’s going to be the next set of advances is probably going to be around what’s people call unsupervised learning. You know, AI today, it’s mostly done by having the system be trained on millions of examples. And, the examples have to be labeled by human, as to what their category is. And that’s not very sustainable, because in a lot of cases hard to get a lot of examples like that. So, we have to get systems that learn from data, but without the data being carefully labeled by humans. And that’s as [inaudible 00:57:19] called unsupervised learning, the quote dark matter of AI.
Kirill Eremenko: That’s a beautiful quote.
Melanie Mitchell: Yeah, it has to happen. We have to figure out how to successfully train systems in an unsupervised way. But right now no one really knows how to do that very well. So I think that’s actually going to be an area where there’s, we’ll, we’ll see a lot of progress soon.
Kirill Eremenko: Got you. Thank you. So, cyber security and unsupervised learning for those listening. Also Melanie, well on that note, we’ve slowly approach to end. Thank you so much for coming on the show. And, before I do let you go, please let us know, what’s the best ways for all listeners to get in touch and follow your work?
Melanie Mitchell: So, they can go to my, my webpage. I don’t know if you have a, you can put that on-
Kirill Eremenko: Yes. We’ll put that in the show notes. Of course.
Melanie Mitchell: Yeah. And that has my contact information and all of my papers and everything, so that’s probably the best way to follow them. Follow my work.
Kirill Eremenko: Awesome. Okay. You also have Twitter? I believe.
Melanie Mitchell: I have Twitter. That’s right.
Kirill Eremenko: Okay. LinkedIn as well?
Melanie Mitchell: And LinkedIn. Yeah.
Kirill Eremenko: And you mentioned before the start of the podcast, you have a course in the Santa Fe Institute, about complexity and it’s free. Tell us a bit more about that. That’s, that’s a course that anybody can take?
Melanie Mitchell: Yeah. The Santa Fe institute has, an online educational platform called Complexity Explorer. Maybe you’ll put that in the course notes. Complexityexplorer.org.
Kirill Eremenko: One word?
Melanie Mitchell: You can put that in the show notes. One word, complexityexplorer, and then .org. The site has many courses and tutorials related to complex systems. My course, I have an online course, they’re called introduction to complexity, which is, has no prerequisites, anyone can take it. And it’s a pretty fun, easy way to get an introduction and an overview of complex systems. It’s kind of based on my complexity book and I’m hoping to do one of those on AI as well. But that’s, that’s for the future.
Kirill Eremenko: Awesome, fantastic. Of course, guys, look all it’s for the book, the new one. The one, the one that we mentioned today was… We mentioned two books, right? So, one is the existing book, Complexity: A Guided Tour. And the new one, Artificial Intelligence: A Guide for Thinking Humans, that’s coming out in September.
Melanie Mitchell: Yeah.
Kirill Eremenko: Okay. On that note, Melanie, thank you so much for being on the podcast. I really appreciate your time and you sharing your knowledge of [inaudible 01:00:13].
Melanie Mitchell: Oh, it’s been great. Thank you so much for having me.
Kirill Eremenko: So, there you have it ladies and gentlemen, that was Melanie Mitchell, professor at Portland State University and one of the leading researchers in this space of AI. What a podcast and how many different resources that were mentioned today. First of all, my favorite part of the podcast was probably the whole notion of complexity and never understood it as clearly before, but indeed it looks like there are lots of commonalities between different systems around the world, starting from ant hills to the human brain to the internet and many more. And it’s very interesting to learn more about that.
Kirill Eremenko: And speaking about learning as Melanie mentioned, you can get free access to her course on complexity if you head on over to complexityexplore.org all one word. Plus of course I make sure to check out Melanie’s books. She’s got six of them and the seventh one is coming out this September, in 2019 and that one is going to be about artificial intelligence.
Kirill Eremenko: As usual you can get all of the links and the materials that we mentioned. I know it might be hard to keep track of all of them in your mind right now, but don’t worry, you can just head on over to www.superdatascience.com/257 that’s, www.superdatascience.com/257, where you will find all of the links and materials mentioned on the show, including all URLs to Melanie’s LinkedIn and Twitter where you can follow her. Her course and her books and plus you’ll get the transcript for this episode if you’d like to check it out.
Kirill Eremenko: On that note, thank you so much for being here today. I hope you enjoyed this chat. Don’t forget to leave a review on iTunes or wherever else you’re listening to this podcast, and I look forward to seeing you back here next time. Until then, happy analyzing.