Voices in AI – Episode 10: A Conversation with Suchi Saria

.voice-in-ai-byline-embed {
font-size: 1.4rem;
background: url(https://voicesinai.com/wp-content/uploads/2017/06/cropped-voices-background.jpg) black;
background-position: center;
background-size: cover;
color: white;
padding: 1rem 1.5rem;
font-weight: 200;
text-transform: uppercase;
margin-bottom: 1.5rem;
}

.voice-in-ai-byline-embed span {
color: #FF6B00;
}

In this episode, Byron and Suchi talk about understanding, data, medicine, and waste.




0:00


0:00


0:00

var go_alex_briefing = {
expanded: true,
get_vars: {},
twitter_player: false,
auto_play: false
};

(function( $ ) {
‘use strict’;

go_alex_briefing.init = function() {
this.build_get_vars();

if ( ‘undefined’ != typeof go_alex_briefing.get_vars[‘action’] ) {
this.twitter_player = ‘true’;
}

if ( ‘undefined’ != typeof go_alex_briefing.get_vars[‘auto_play’] ) {
this.auto_play = go_alex_briefing.get_vars[‘auto_play’];
}

if ( ‘true’ == this.twitter_player ) {
$( ‘#top-header’ ).remove();
}

var $amplitude_args = {
‘songs’: [{“name”:”Episode 10: A Conversation with Suchi Saria”,”artist”:”Byron Reese”,”album”:”Voices in AI”,”url”:”https:\/\/voicesinai.s3.amazonaws.com\/2017-10-16-(00-59-07)-suchi-saria.mp3″,”live”:false,”cover_art_url”:”https:\/\/voicesinai.com\/wp-content\/uploads\/2017\/09\/voices-in-ai-cover.png”}],
‘default_album_art’: ‘https://gigaom.com/wp-content/plugins/go-alexa-briefing/components/external/amplify/images/no-cover-large.png’
};

if ( ‘true’ == this.auto_play ) {
$amplitude_args.autoplay = true;
}

Amplitude.init( $amplitude_args );

this.watch_controls();
};

go_alex_briefing.watch_controls = function() {
$( ‘#small-player’ ).hover( function() {
$( ‘#small-player-middle-controls’ ).show();
$( ‘#small-player-middle-meta’ ).hide();
}, function() {
$( ‘#small-player-middle-controls’ ).hide();
$( ‘#small-player-middle-meta’ ).show();

});

$( ‘#top-header’ ).hover(function(){
$( ‘#top-header’ ).show();
$( ‘#small-player’ ).show();
}, function(){

});

$( ‘#small-player-toggle’ ).click(function(){
$( ‘.hidden-on-collapse’ ).show();
$( ‘.hidden-on-expanded’ ).hide();
/*
Is expanded
*/
go_alex_briefing.expanded = true;
});

$(‘#top-header-toggle’).click(function(){
$( ‘.hidden-on-collapse’ ).hide();
$( ‘.hidden-on-expanded’ ).show();
/*
Is collapsed
*/
go_alex_briefing.expanded = false;
});

// We’re hacking it a bit so it works the way we want
$( ‘#small-player-toggle’ ).click();
$( ‘#top-header-toggle’ ).hide();
};

go_alex_briefing.build_get_vars = function() {
if( document.location.toString().indexOf( ‘?’ ) !== -1 ) {

var query = document.location
.toString()
// get the query string
.replace(/^.*?\?/, ”)
// and remove any existing hash string (thanks, @vrijdenker)
.replace(/#.*$/, ”)
.split(‘&’);

for( var i=0, l=query.length; i<l; i++ ) {
var aux = decodeURIComponent( query[i] ).split( '=' );
this.get_vars[ aux[0] ] = aux[1];
}
}
};

$( function() {
go_alex_briefing.init();
});
})( jQuery );

.go-alexa-briefing-player {
margin-bottom: 3rem;
margin-right: 0;
float: none;
}

.go-alexa-briefing-player div#top-header {
width: 100%;
max-width: 1000px;
min-height: 50px;
}

.go-alexa-briefing-player div#top-large-album {
width: 100%;
max-width: 1000px;
height: auto;
margin-right: auto;
margin-left: auto;
z-index: 0;
margin-top: 50px;
}

.go-alexa-briefing-player div#top-large-album img#large-album-art {
width: 100%;
height: auto;
border-radius: 0;
}

.go-alexa-briefing-player div#small-player {
margin-top: 38px;
width: 100%;
max-width: 1000px;
}

.go-alexa-briefing-player div#small-player div#small-player-full-bottom-info {
width: 90%;
text-align: center;
}

.go-alexa-briefing-player div#small-player div#small-player-full-bottom-info div#song-time-visualization-large {
width: 75%;
}

.go-alexa-briefing-player div#small-player-full-bottom {
background-color: #f2f2f2;
border-bottom-left-radius: 5px;
border-bottom-right-radius: 5px;
height: 57px;
}

.voice-in-ai-link-back-embed {
font-size: 1.4rem;
background: url(https://voicesinai.com/wp-content/uploads/2017/06/cropped-voices-background.jpg) black;
background-position: center;
background-size: cover;
color: white;
padding: 1rem 1.5rem;
font-weight: 200;
text-transform: uppercase;
margin-bottom: 1.5rem;
}

.voice-in-ai-link-back-embed:last-of-type {
margin-bottom: 0;
}

.voice-in-ai-link-back-embed .logo {
margin-top: .25rem;
display: block;
background: url(https://voicesinai.com/wp-content/uploads/2017/06/voices-in-ai-logo-light-768×264.png) center left no-repeat;
background-size: contain;
width: 100%;
padding-bottom: 30%;
text-indent: -9999rem;
margin-bottom: 1.5rem
}

@media (min-width: 960px) {
.voice-in-ai-link-back-embed .logo {
width: 262px;
height: 90px;
float: left;
margin-right: 1.5rem;
margin-bottom: 0;
padding-bottom: 0;
}
}

.voice-in-ai-link-back-embed a:link,
.voice-in-ai-link-back-embed a:visited {
color: #FF6B00;
}

.voice-in-ai-link-back a:hover {
color: #ff4f00;
}

.voice-in-ai-link-back-embed ul.go-alexa-briefing-subscribe-links {
margin-left: 0 !important;
margin-right: 0 !important;
margin-bottom: 0.25rem;
}

.voice-in-ai-link-back-embed ul.go-alexa-briefing-subscribe-links a:link,
.voice-in-ai-link-back-embed ul.go-alexa-briefing-subscribe-links a:visited {
background-color: rgba(255, 255, 255, 0.77);
}

.voice-in-ai-link-back-embed ul.go-alexa-briefing-subscribe-links a:hover {
background-color: rgba(255, 255, 255, 0.63);
}

.voice-in-ai-link-back-embed ul.go-alexa-briefing-subscribe-links .stitcher .stitcher-logo {
display: inline;
width: auto;
fill: currentColor;
height: 1em;
margin-bottom: -.15em;
}

Byron Reese: This is “Voices in AI” brought to you by Gigaom. I am Byron Reese. Today, my guest is Suchi Saria. Where do I start when going through her career? She has an undergraduate degree in both computer science and physics. She has a PhD in computer science from Stanford, where she studied under Daphne Koller. She interned as a researcher at IBM and at Microsoft Research, where she worked with Eric Horvitz. She is an NSF Computing Innovation fellow at Harvard, a DARPA Young Faculty Award Winner, and is presently a professor at Johns Hopkins. Welcome to the show.

Suchi Saria: Thank you.

Let’s start off with the biggest, highest-level question there is. What is artificial intelligence? How do you answer that question when it’s posed to you?

That’s a great question. I think AI means very different things to different people. And I think experts in the field, at the high level, understand to a degree of what AI is; but they never really posit a very concrete, mathematical description of it. Overall, our goal is… We want computers to be able to behave intelligently, and that’s really the origin of how the AI field of computer science emerged.

Now along the way, what has happened is… Starting from really classical applications— autonomous driving or image recognition or diagnostics… As lots and lots of data has been collected, people have started to develop numerical tools, or statistical methods, or computational methods that allow us to leverage this data, to build computers or machines that can do useful things that can help humans along the way. And so that then has also become part of AI.

Effectively, the question that as a field we often ask ourselves is: Does AI really mean useful tools that help humans, automating a task that humans can do, giving computers the ability to also do? Or is it going at properties like creativity and emotion, that are very interesting and unique aspects of what humans often exhibit? And do computers have to exhibit that to be considered ‘artificially intelligent’? So really there’s a debate about what is intelligence, and what does it really mean; and I think different experts in these fields have very different takes on it.

I only ask it because it’s a really strange question. If you ask somebody at NASA, “What is space travel?”, maybe there’s… “Where does space begin?” or “How many miles up?”—If you ask all these different fields, they kind of know what the field is about. You said something I never heard anybody say, which is: “Those of us who are researchers in it, we have a consensus on what it is.”

I would say at a very high level, we all agree. At a high level, it is the ability to systematize or help computers behave and reason intelligently. The part that is left to be agreed upon is ‘behave and reason intelligently the way humans do’. The way humans do things is important because, in some fields, we should study humans; we should understand the way humans do it and biological systems do it, and then build computers to do it the way humans do it.

In other fields, it’s not so important that we do it exactly the way humans do it. Computers have their own strengths, and effectively, perhaps what’s more important is the ability to do something, rather than the process by which we’re getting there. So we all agree that the goal is to build intelligent machines.

Intelligent machines that crunch a lot of data, intelligent machines that can reason through information that’s provided, produce what needs to be done, interact intelligently—and by that, we mean understand the person that’s in front of you, and understand the scenario that’s being presented to you, and react appropriately.

Those are all things we’ll agree on. And then, effectively, the question is: Do we need to do it the way humans are doing it? In other words, is it in the making of human intelligence, or is it about giving this capability to machines by whichever way the machines are able to learn that?

I won’t spend too much time here, because it may not be interesting to everyone else, but to say artificial intelligence is teaching machines to reason intelligently—you’re using ‘to reason intelligently’ to define the term ‘intelligence’. Doesn’t that all obfuscate what intelligence is?

Because at one extreme end, it’s defined simply as something that reacts towards its environment; a sprinkler system that comes on when the grass is dry is intelligent. On another extreme end, it’s something that learns, teaches itself; it evolves in a way that your sprinkler system doesn’t. It’s a learning system that changes its programming as it’s given more data.

Isn’t there some element of what intelligence is that we all have to circle around, if we are going to use this term? And if we’re not going to circle around it, is there a preferred way to refer to this technology?

Yeah, I think the preferred way is the way we think about it. I think the other aspect of the field that I really love is the fact that it’s very inclusive. The reason the field has moved forward so quickly is because, as a field, we’ve been very inclusive of ideas from psychology, from physics, from neuroscience, from statistics, from mathematics—and of course, computer science.

And what this really means is as a field, we move forward really quickly and there’s really room for multiplicity of opinions and ideas. The way I often think about it is: What’s the preferred way people like me think about it—and others might give you different opinions about it, but fundamental to all this is the idea of learning.

Rather than building brittle systems that effectively have hard-coded logic which says, “If this happens, then do this. If that happens, then do this”—what’s different here is that effectively these systems are more designed to program their own logic, based upon data. They’re learning in a variety of different ways—they learn from data. Data where in the past, people have presented a scenario.

Let’s say in this scenario, you might consider how another intelligent human or an expert human is reacting to the scenario, and you’re watching how the human behaves or reacts; and from that, the computer is trying to learn what is optimal. Alternatively, they may learn by interacting with their environment itself. For instance, if the environment has a way… Like in the game of Go, the environment here being the board game itself—had a way of giving feedback… A version of feedback would be, if you make a move, you get a score attached to whether or not this is a good move, and whether or not it will help you win, and they’re basically using that feedback.

It’s often the type of feedback we as humans use all the time in real life… Where effectively you could imagine kids… If there’s a pot that’s too hot and they touch it, next time they see a similar object, they’re much less likely to touch it. And you know as adults, we go and we often analyze scenarios around us, and see if something has a positive or a negative feedback.

And then, when we see negative feedback, we sort of register what might have caused it, the reason about what might have caused it, and try to do that process. The notion of learning is pretty fundamental. The way by which it learns is really a huge body of work which has focused on that—which is, how do we develop more general purpose methods by which computers can learn, and learn from many different types of data, many different types of supervision and effectively, learn as quickly as possible?

You used the word ‘understand’ the person, ‘understand’ the situation. There’s a famous thought experiment on that word, and what the implications are. It’s called The Chinese Room Problem, and just to set it up for the listener… There’s a man who speaks no Chinese—we call him the librarian—and he’s in this giant room with thousands and thousands of these very special books.

And people slide questions under the door to him, and they’re written in Chinese. He doesn’t understand them, but he knows to match the very first symbol in the message to the spine of a book, pulls that book down, looks up the second symbol that directs him to another book, and another one, and another one… Until he finally gets to the end of this process.

And he copies down the characters that he sees, slides that back out, and it’s a perfect answer in Chinese. And of course, the man doesn’t know the meaning of what it was about, but he was able to produce this perfect answer using a system. The question is: Does the man understand Chinese?

And of course, the analogy is obvious. That’s all a computer is doing; it’s running a deterministic program, and so forth. So I put the question to you: Does the man understand Chinese? Can a computer understand something, or is understanding just a convenient word we use, but [where], clearly, the computer doesn’t understand anything?

Let’s shift our attention for a second away from computers and to humans. I often think hard about… I try to pull out scenarios where I’m wondering, am I effectively running an algorithm? And what is my own algorithm?—and even considering scenarios where it’s not so prescriptive. Perhaps I needed to be creative.

My job often involves being creative, and coming up with new ideas frequently. And the question I ask myself is: Am I just deriving this idea out of previous experiences that I already had? In other words, am I effectively just engaging in the task of…

Let’s say I have A and B, and this really creative idea… But what my brain has become really good at is, in new scenarios, quickly figuring out what are the relevant elements like A, B and C in my past which are pertinent, and then from that, coming up with something that looks like a combination or variation. In other words, it’s not as big a leap of faith as it [might seem] to someone who doesn’t have my experience, or doesn’t have my background.

And then, I think hard about it, and perhaps it really is just derived from the things I know. What this is getting at is me being a little cynical about my own ability to—my own assessment of how much do I really understand. Is understanding effectively the ability to quickly parse information, determine what’s important, apply rules of logic and a bit of randomness in order to experiment with ideas and then come up with a new idea?

I don’t really have an answer to this. But I’ve often wondered, this is maybe what we do; and our ability to do this really rather quickly is sort of what distinguishes different humans in their ability to understand and come up with a creative idea quickly. And so, if I think about it from this point of view, it doesn’t seem to me a complete stretch to imagine that we could teach computers to do these things.

So let me give you an example. For instance, going back to the very popular news story around AlphaGo, when the AlphaGo started to explore new moves. Many individuals who are not familiar with the topic of AI thought, “Wow, that’s amazing! It’s being creative, it’s coming up with brand-new moves altogether”—that humans, human experts hadn’t really known. But really, all it was doing is doing search, in some super large space.

And its ability to do search is pretty expansive. And the other thing it has is really clever ways of doing search, because it has heuristics that is has built up from its own experience of doing learning. And so, in a way, that’s really what humans are doing, and that’s really what experience gives us. So let’s go back to your question of: Does the [Chinese Room person] really understand?

I think my personal issue is that I don’t know what understanding really means. And the example I gave you… If you were to define understanding that way, then I think in today’s world, we would say maybe that man didn’t understand what he was doing, but maybe he did. I’m not sure. It’s not obvious to me.

Do we measure understanding by the output—the fact that you give an input and they give a reasonable output? Or do we measure it by some other metrics? It’s a really great question though.

You captured the whole debate in what you just said, which is… The room passes the Turing test. You wouldn’t be able to tell—if you’re the Chinese speaker outside the room passing in the messages—you wouldn’t be able to tell if that was not a native speaker on the other side.

And so the machine ‘thinks’. Many people in the field had no problem saying the man understands Chinese, but at a gut level, that doesn’t feel right. Because the man doesn’t know if that was a message about cholera or coffee beans, or what—it’s just blinds on paper. He knows nothing, understands nothing, just walks through some old thing that gets him to copy these marks down on paper.

And to say that is understanding trips people up. The question is: Is that the limit of what a machine will ever be able to do? I will only say one thing, and then I would love your thoughts. Garry Kasparov kind of captured that when he lost to Deep Blue back in ‘97. He said, “Well, at least it didn’t enjoy beating me.”

His experience of the game was different from the computer’s experience of the game. I only think it’s a meaningful question because it really is trying to address the limits of what we can get machines to do. And if, in fact, we don’t understand anything either, then that does imply we can build AGI and so forth.

I agree with you. I think it’s a very meaningful question. And I certainly think it’s a topic we should continue to push on and understand more deeply. I would even go back, to say that I bet there are people around you—maybe not as holistic and expansive a context as the Chinese man you described—but you could imagine scenarios where somebody is really good… their whole job is sort of like, they learned numerous algorithms like this.

And you could imagine colleagues like that… Where they’re effectively really good at fielding certain types of questions, and pushing data out. And maybe they have not built the algorithm, but they understand what the person in front of them is asking, and they understand what kinds of answers they need to hear in order to be able to answer questions in a satisfactory manner.

Effectively, my point is even though we think that in the example… [If] somebody told you he doesn’t understand, that [conclusion] is very possible. If nobody had told you that, and he always was able to produce something that was acceptable or of a high quality, everybody else would always think of this person as, “He understands what he’s doing.” And we probably have people like that around us. We’ve all experienced this to some extent.

It could be. And If that is the case, it really boils down to what the word ‘artificial’ means in ‘artificial intelligence’. If ‘artificial’ means it’s not really intelligence—like artificial turf isn’t really turf —if it really means that, then you’re right. As long as you don’t know that he doesn’t understand, it doesn’t really matter.

I would love to ask one more question along these lines. I’m really intrigued by what we will need to do to build a machine that is equivalent to a human; and I think your approach of, “Let’s start with what humans do and talk about computers later” is really smart.

So I would put this to you… Humans are sentient, which is often a word that is misused to mean intelligent. [But] that’s actually ‘sapient’. ‘Sentient’ means you’re able to feel things—usually pain—but you’re able to feel something, to have an experience of feeling something.

That’s kind of also wrapped up in consciousness, but we won’t talk there yet…

Is it possible for a computer to ever feel anything? It’s clearly possible to set up a temperature sensor that, when you hold a match to it, the computer can sense the temperature; and you can program the computer to scream in agony when it passes a certain temperature. But would it ever be possible for a computer to feel pain, or feel anything?

Let’s step back and ask the following question… Two parts: First is, “To make computers that feel something—can that be done?” The second question is, “Why do we need computers that feel things?” Is that really what separates artificial intelligence from human intelligence?

In other words, is that really the key distinction? And if so, can that be built? Let’s talk about how do we build it. Have you heard, or have you seen, any of the demos out of this terrific company—I think it’s called Hanson Robotics. If you go online, you can Google it, you can search for it. David Hanson is one of the founders, and effectively, what they build is a way to give a robot a face; and he has these actuators that allow very fine-grained movement.

And so, effectively, you see full facial features and full facial expressions projected onto a robot. The robot can smile and the robot can frown, and it can get angry and it can stare and express excitement and joy. Effectively, he’s sort of done a lot of the work of—not just what it takes to build mechanically those parts, but also thinking harder about how it would get expressed, and a little bit about when it would get expressed.

And then independently, there’s great work from MIT—and you know, other labs, too—but I’m just thinking of one example: They looked at learning and interpreting emotion. For example, you might imagine [that] if the person in front of you is angry, you might want the robot to react and respond differently than if the person was happy and excited.

Effectively, you could imagine putting a camera, seeing the stream coming in, [and] the computer processes it to do classification for whatever type of emotion is being expressed—you could specify a list of emotions that are commonly expressed. From that, the computer can then decide what human emotion is being expressed, and then decide what emotion it wants to express.

And now, you can imagine feeding it back into Hanson’s program that allows them to generate robotic facial motions that are effectively expressing emotion, right? So if we had to build it, we could build it. We know how to think about building it. So mechanically, it is not impossible. So now the piece here is—the second question is: If we could do this, and in fact there are studies that…

For instance, when I was with Microsoft Research, there was a robot that would greet you, and it would basically see where you were standing, and it would turn its head to try to point to you. And many, many individuals who weren’t familiar with robotics—many visitors who would come to Microsoft, people that weren’t in the technology industry, but were just visiting—would see that and get really excited, because the idea of a robot turning its head and moving its eyes in response to where you’re standing was cool, and seemed very intelligent.

But effectively, if you break down the mechanics of how it’s doing it, it’s not a big surprise. Similarly, you could augment it by also showing facial expressions, and I think CMU— Carnegie Mellon—has a beautiful robot that’s called the robot receptionist; her name is Valerie. They worked on it at the drama department at Carnegie Mellon.

And they basically filled the robot with lots of stories, and it was really funny… As a graduate student, I was visiting, and met Valerie for the first time… You could ask her for directions, and she would give you directions on where to go. If I could say, “Where’s Manuela’s office?” the robot would point me to where it is.

But in the middle, she would behave like a human, where she would be talking on the phone to her sister; and they’d be talking about what’s going on, what’s been keeping them busy, and they’d hang up or she’d put people on hold if a new visitor came in, and so forth.

So what I’m challenging is this concept of, is it really the lack of human emotion, or what you consider to be human-like emotion—to be very special to humans? Is it that? Is it mimicking that? What does it mean to feel pain? Is it really the action-reaction—somebody’s poking you and you react—or is it the fact that there’s something internal, biological that’s going on, and it’s the perception of that?

That could be. You asked a good question: Does it matter? And there would be three possible reasons it would matter: First, there are those that would maintain that an intelligence has to experience the world, that it isn’t just this abstract ones and zeros it-lives-in-a-computer thing—that a true intelligence would need to be able to actually have experiences.

The second thing that might make it matter is… There was a man named Weizenbaum who famously created a program in the ‘60s called ELIZA, which was a really simple program. You would say, “I’m sad.” It would say, “Why are you sad?”

“I’m sad because my brother yelled at me.”

“Why did your brother yell at you?”

And Weizenbaum turned against it all, because what he saw is that even people who knew it was just a very simple program developed emotional attachment to it. And he said… When the computer says, “I understand,” as Eliza did, he said it’s just a lie. There is no ‘I’ in there, and there’s no understanding.

But really the reason why it might actually matter is another thought experiment, that I will put to you and to those listening: It’s the problem of Mary.

Mary is a hypothetical person who knows everything about color. She knows literally everything, like at a god-like level. She knows everything about photons and cones and how color manifests in the brain. She knows everything that there is to know about it, but the setup is that she has never seen it. She lives in this room that’s all black and white, and only has black-and-white computer monitors.

She walks outside one day and sees red for the first time. And the question is: Did she learn something new? Is experiencing something different than knowing something? And if you say yes… It’s one of those things where most people, at first glance, would say, “Yes, if she’s never seen color and she sees it for the first time; yes, she learns something.”

And if that is the case, then a computer has to be able to experience things in order to learn past a certain point. Do you think Mary learned something new when she saw color for the first time? Or no, she knew exactly what it would look like, and experiencing it would make no difference?

So, you know what Mary knew. Did she know ahead of time what red would look like when she stepped out?

Well, she knew everything about color. She never saw it, but she knew exactly what it would do to her brain—at a molecular level, atomic level—every single thing that would happen in her brain when she saw a color, but she’s never seen it.

As a computer scientist, when you say that you me, I would say that the representation of what Mary understands or know is ambiguous. What I mean by this is, I don’t know what it means to say—I understand what it means to say “she knows at the molecular level what happens.” I understand what it means to say she knows, perhaps, about the relationship between different primary colors, and the derivative colors and so forth.

But are you saying that she knows… Is it the case that she receives an image using her eyes, and her eyes represent it using some form of internal neuronal format?—Are you saying she knows that? Because if she doesn’t know that, then effectively, she still has a partial understanding of what knowing everything about color means.

So this might be an interesting place… Where we think her knowing everything about color…

If you tell me: Somebody presented a red image to her, and she knew what it meant to take that red image and convert it—and these are really hypotheticals; I’d have to understand this more deeply and really study it, and perhaps bring in someone who understands human perception really well—but my first step-check would be: What does it mean for her to know everything about color?

And what if we present her with an image, her visual cortex processes it, and effectively, she is getting data, and she is seeing it internally. Is it stored in RGB format? Is she storing it in some format that she understands? Is she aware? Has that core process happened in her head before? It may not have been due to her stepping out, but the question is: Is that something that she is privy to, or has knowledge of?

And if so, then I would say that when she steps out… And if all she is doing is focusing on the color red, and that is the only sensation that’s being generated in her head; then yeah, this is going to seem familiar to her because it’s something she’s seen before. The word ‘experience’ at that point is a really interesting word. And it would be fun to sit down and try to write down formal definitions for what it means.

And generally, we think of having ‘seen’ and having ‘experienced’ as two different things, in human emotions. But I think from a computer point of view, they don’t seem different. Even as a human, if I think hard about it, I don’t know really what the distinction is. I don’t know what it means to kind of know it, to know it, and then experience it. What is the difference between those things?

It may be that the question imperfectly captures it, because it’s formed very casually, but… Humans experience the world.

You taste a pineapple, and what that pineapple tastes like… Tasting it seems to be a different thing than knowing something. If I know what it tastes like, it’s a different thing than actually having the experience of tasting it.

Knowing how to ride a bicycle is different than having ridden a bicycle, and knowing how you feel balanced when you get on one. Touching something warm feels a certain way that knowing all about warmth does not capture.

And so, the question is: If a machine cannot actually feel things, touch things, taste things, have any experience of the world—then whatever intelligence it has is truly fake. It really is artificial in a sense that’s completely fake.

And you’re right, I think, in asking the question… Why we ask these questions… And a lot of what people are often doing is asking questions about people. Are people machines? Are we…

But then they have this disconnect, to say: “But we feel, and we experience, and we know, and those seem to be different than things my iPhone can do.” So I think I’m trying to connect those dots to say, experiencing something seems to be different than knowing something.

But you’re right; it’s imperfectly formed. I’ll let you comment on that, and then let’s move on to your research, because there’s so much there I would love to hear more about.

Sure! So I think I am going to continue to push back a little bit on… I feel that people’s experience of what they believe a machine or an iPhone can do is very much based on… I think it’s easier to think about a single narrow task.

You could take the task of eating a pineapple, or the task of going and experiencing a warm day… But effectively, the way I think about it is [that] a lot of these capabilities don’t exist because most people haven’t thought that building a machine that eats a pineapple is a very useful thing, so people haven’t bothered to build it.

But let’s imagine I decided that was important, and I wanted to build it. Then, what I would do is much like—going back to David Hanson… I would try to first identify what do I mean by ‘experience eating a pineapple’, and if the idea is that every time I am given it—a tasty pineapple—I can eat it and it’s delicious, and my eyes light up. And if I eat a rotten pineapple, then I’m visibly upset.

Then I could imagine building the sensor to which you feed the pineapple. It runs chemical tests that check, effectively, what’s in the pineapple and… You could start by version one. Version one tests what’s in the pineapple, and based on that—and it’s hooked up to David Hanson’s robot—and it generates the reaction, which is excited, or sad, or unhappy, and visibly unhappy, or sad, depending on how tasty or not-so-tasty the pineapple is.

And you could even take it a step further by saying, “You know what? I’m going to give lots of humans things to eat; and based on that, I will watch what the humans are doing. And then effectively, the computer’s just learning by taking the same fruit and eating it itself. And you didn’t even program anything about how to react. All it did was watch humans eat it, and based on that, it learned that when certain molecular compositions exist in the thing it’s tasting, then it tends to get happy or less happy.

And you might imagine it starts to mimic. In fact, we could take it even another step further and say, “Let’s give a group of robots the same set of sensors, and they have to figure out a way by which they communicate and barter with each other.” So effectively, there’s an objective function, and the objective function—or the goal for the group of robots—is to figure out an effective way to trade.

The trade is such that one group of robots loves apples. The other group of robots loves pineapples. And the way you know that is, effectively, they’ve each lived in different environments and—I don’t like the word ‘live’, because it’s over-interpretive…

What I mean is, they’ve been trained in different environments, and the ones that love to eat apples have learned to get an excited expression to good apples, and the other set of robots get an excited expression to good pineapples. And you want them to work together to trade, such that everybody is as happy as possible.

Then it’s completely possible they’ll be able to effectively learn, on their own, a trading strategy where they say, “You know what? The people who don’t like pineapples should give away their pineapples, and the people who don’t like apples should get rid of apples.” So, effectively, what I was giving you was an example where…

If we understand what is the objective we’re after—which is, what does experiencing a pineapple mean—then very often, you can turn it into some mathematical objective by which the computer can learn how to do similar things, and very quickly… ‘Very quickly’ depends a lot on the complexity of the task—but it can mimic that behavior or goal—and now I use the word ‘mimic’ lightly…

But effectively it can, be it similarly, or—and one could argue, “What does ‘similar’ mean?” and, “What does ‘behave similarly’ mean?”… But for the most part, we would look at this and be pretty satisfied that it’s doing something that we would consider to be intelligent. We would consider it to be experiencing something.

Unless, the only block in our head is we think it’s a machine… So it’s hard because we think humans experience things and machines don’t… But what I think would be really cool is to think about, “Are there tasks where we really experience something, that we think there is no way to build a machine to experience the same thing?” What does it mean to experience in that setup?

I think that would be interesting, and I would love to hear [from] our listeners who have ideas, or want to send me ideas. I would love to hear that!

Well, I think the challenge, though, is that in civilization we’ve developed something called ‘human rights’, where we say: “There are things you can’t do to a person no matter what. You can’t torture people for amusement, and you can’t do these things.”

So we have human rights, and we extend them—actually broadly—to other creatures that can feel pain, so we have laws against cruelty to animals because they feel pain.

It sounds like you’re saying the minute you program a computer to be able to mimic a frown, or to scream and mimic agony, that that is somehow an equivalency; and therefore, we need laws that… Once the temperature hits 480 degrees, the computer screams, and we need to outlaw that; we need to grant those things rights, because they are experiencing things.

And then, you would push it one step further to say, when I am trying to get my car out of the mud, and it’s smoking, and the gears are grinding… That that too is experiencing pain, and therefore that should be…

You run into one of two risks. You would either make the notion of, “Things that feel have rights not to be tortured”—you either make that ludicrous, by applying it to anything that can make a frowny face…

You either try to elevate everything that’s mechanical to that, or you end up debasing people, by saying: “You don’t actually feel anything. That’s just a program. You’re a machine, and you don’t actually have any experience. And you reporting pain, it isn’t real. It’s just you, kind of, programmed to say that.”

How do you have rights in a world where you have that reductionist view of experience?

Personally, I think it’s pretty liberating that computers don’t get tired, and they don’t feel pain. When I say the word ‘feel pain’, I mean feel pain in the sense that, if you ‘hurt’ me a lot in a certain way using a pin, I may screech. But also, I could shut down, I could stop being productive.

But if you take a computer, and it has a hard, metal shell… And you take a pin and you effectively poke it too hard, it doesn’t really do much to the computer because it’s fine.

But then, there are other things… For instance, if you unplug the computer, it’s dead. And there’s an equivalent notion of unplugging me. So for me, I kind of find it liberating that we don’t have to try to do all the same things. The thing that is very exciting to me about it is that this has its own strength. A machine is effectively a very… I think there’s two takeaways for me, personally: One, the fact that it makes me think harder about, “What do I have to do to be special?”—about myself.

So effectively, there are lots of things that I used to consider to be very special—I’m still special, of course [laughs]—but what I mean is, I would attribute this mystical sense to—which is maybe not so necessary… Like the whole task of programming computers and developing these learning machines has really made me a little bit more humble about what I consider to be very hard, and not-so-hard; and effectively realizing that maybe some of these properties that humans exhibit can actually be demystified, right?

I understand a little bit more about, what does it mean to do X and do Y? It makes me think harder about something that comes so naturally to us—how is that we do it? How is it that different beings do it? And the fact that computers can do it, and maybe it’s not exactly the same way, and it’s a slightly different way…

So just having that awareness is actually pretty exciting, because it makes things that are everyday around us, which are pretty rote… not so rote anymore. It’s fun to watch people walk. You’re sort of saying, “Ah, it’s so natural and easy for them,” but if you really think about it, there are just so many complicated things we are doing. And then, you try to make and teach computers how to walk, you sort of very quickly realize how complicated it is, and it’s kind of cool that we as human beings can do it.

So effectively, one of the aspects of it is it teaching me a little bit more about myself, and realizing the complexity and also the steps or procedures it takes for me to do some of the things that I’m doing. The second aspect of it is realizing that perhaps it’s a good thing [that] there are certain things a computer is good at, and things that it’s not good at… And perhaps, taking advantage of that in order to build systems that are useful in practice, and can really make us, as a society, better off is pretty exciting to me.

So I think the idea of trying to exactly mimic humans—or whether we would be able to exactly mimic humans—is sort of interesting; but practically speaking, I don’t think of it as the most interesting consequence of this… or the area of debate for most experts in the field.

We think more of it as, what are areas where we can really build useful things that could then help us make humans faster, make everyday life better, save us work—that would be better to pass off to a computer to do, so that it frees up time for us to do other things.

But… Does that answer your question a little bit more, about human rights? So effectively, I think the issue was, if you are concerned about pain, then perhaps there should be rules about when humans experience pain, we ought not to do X, Y and Z. Maybe computers could have different sorts of rules, because they experience different sorts of things, and they’re good and bad at different sorts of things.

And I think we just haven’t come to a place where there’s a general agreement among scientists building it, about what is and isn’t useful, and we work around those principles. And that has really dictated what gets built.

Fair enough! So tell me about… You have an unusual professorship at Johns Hopkins. What is that? Can you talk about your work there?

Yeah, sure! I’m a faculty in Computer Science and Stats, but also, I’m a faculty in Public Health. Hopkins is one of the largest schools of public health in the country; and in particular, I am in the department of Health Policy and Management. So what’s unique about my appointment is that…

Hopkins has a very large School of Public Health, a very large School of Medicine. And I effectively interact—on a day-to-day basis—not just with engineers, but also people who are clinical experts and public health experts who design policy… [Which brings] a multifaceted view into the kinds of questions we’re trying to answer around using computers, and using data-driven tools to improve medicine and improve public health.

And so, what does that look like on a day-to-day basis? What kinds of projects are you working on?

Let’s see… Let me give you a concrete example.

One area of study that we spend time on is detecting adverse events in hospitals. They’re called ‘hospital-acquired complications’. One example of this is sepsis. And effectively, what happens is, let’s say a patient is coming into the hospital for any condition; and sometimes they come in because they have an infection, and the infection goes undetected, and turns into what’s called sepsis.

Sepsis is effectively when your body is trying to fight the infection, [and] it releases chemicals, and these chemicals start attacking your [own] organs and systems. This has happened in some fraction of the cases, and if it does happen, it ends up causing organ damage, organ failure, and eventually death if it goes untreated.

And so, this is an example where individuals who have sepsis at the moment… Physicians are relying on visible signs and symptoms in the patient in order to be able to initiate treatment. And what our last work has shown is [that] it’s possible to identify very early, based on lots of data… So when they come in, as part of routine care, they’re taking tons of measurements, and these measurements are getting stored electronically…

And so, what we do is we analyze these measurements in real-time, and we can identify subtle signs and symptoms that currently the physicians miss all the—you know, it’s a busy unit. In a 400-bed hospital, there’s persons coming in, there are lots of other patients; it’s a distributed care team. It’s tough. And if the symptoms are not really visible, or are subtle, they sometimes get missed.

And so, an example area where we’ve shown is—with sepsis, for instance—you can identify very early, subtle signs and symptoms, and identify these high-risk patients and bring this to the caregiver; so that they can now start to initiate treatment faster. And so, this is exciting because it really demonstrates the power of computers: They’re tireless; they can sit there, process data from 400 patients continuously, all the time.

We can learn from expert doctors what are signs and symptoms, but not just that! We can look at retrospective data from 10,000 or 70,000 or 100,000 patients, and understand things like what are the subtle signs and symptoms that happen to appear in patients with sepsis and without sepsis, and use that to start displaying this kind of information to physicians.

And now, they’re better off, because suddenly, they are missing fewer patients. The patients are better off because they can go in completely happy that they’re going to be cared for in the best way possible, and the computer is sitting there, and it really has no reason to complain because all it’s doing is processing the data, and it’s good at that. So that’s one example. And there are lots of other areas.

Another area we’ve been spending time looking at is complex patients, patients of… the word ‘complex patients’ is a little… Let me demystify that a little bit. So looking at diseases where there’s a ton of diversity or heterogeneity in symptom profile; so for example diseases like lupus, scleroderma, multiple sclerosis, where the signs and symptoms vary a lot across individuals. And really understanding which person is going to be responsive to which treatment [in these cases] is not so obvious.

So again, going back to the same philosophy: If we can take data from a large patient population, we can analyze this and start to learn what—for a given patient—is their typical course going to look like, and what are they going to be likely to be responsive to. And then [we can] use that to start bringing that information back to our physicians at the point of… They can now use this information to improve and guide their own care. So those are some examples.

I was just reading some analysis which was saying that before World War II, doctors only had five medicines. They had quinine for malaria, they had aspirin for inflammation, and they had morphine for pain… They had five medicines, and then, you think about where we are today. And that gives one a lot of hope.

And then you think about… We kind of have a few challenges. I mean even all the costs, and all the infrastructure and all of that, just treating it as a mental problem… One, as you just said, no two people are the same, and they have completely different DNA; and they have completely different life experiences. They eat different food for lunch, all of this stuff.

So people are very different, and then we don’t have really good ways to collect that data about them and store it and track it. And so it’s really dirty data over a bunch of different kinds of patients. So my question is: How far do you think we’re going to be able to go? How healthy will we be?

You can pick any time horizon you want. Will we cure aging? Will we eliminate disease? Will we get to where we can sequence any pathogen, and model it with the person’s DNA in a computer, and try 10,000 different cures at once, and know in five minutes how to cure them? Or do we even have a clue of what’s eventually going to be possible?

So I think one of the interesting things, when I first joined Hopkins, that I learned very early, is that when we dream of what an ideal health system ought to look like… Wouldn’t it be great if we had cures for everything? [But] one of the most surprising and disappointing facts I learned was that even in cases where we know what the right treatment is; even in cases that we know—where we could have treated them, had we known upfront, who they were and what was the appropriate sort of therapy for them…

Right now, we have many such cases we miss. So I don’t know if you’ve seen this Institute of Medicine report that came out in 2011 or 2010—I can’t remember the date—where they talk about how a third or a quarter of the amount of money that’s spent in healthcare, they think of it as ‘unnecessary waste’.

Unnecessary waste means waste because we are over-treating; waste in cases where we’ve kept people longer than was necessary; waste because there were complications that were preventable; waste because we gave them treatments that weren’t the right treatments to begin with, and we should’ve given them something else.

And I don’t think the answer is as simple as, “Oh, why isn’t our health system better? Is it because we’re not training the most competent doctors? Is it because our medical educational system is broken?” No. I think if you actually sit inside a hospital, and you watch what’s going on, it’s such a multi-disciplinary, multi-person environment…

That every decision touches many, many people, including the patient. And there’s all this information, and all these decisions have to be made very quickly. And so what to know about any given individual, at any given time, to determine the right thing to do is actually very complicated. And it’s pretty amazing to me that we’re as effective as we are, given the way the system is built up.

So effectively, if you really think about it… To me, a part of it is the system’s problem, in the sense that if, going back… Our delivery of healthcare has very much come out of the era where there were only so many medications. They kind of knew what to do, there were only so many measurements, the rules were easy to store in our head, and you could really focus on execution—which is making sure we’re able to look at the individual and sort of glean what is necessary, and apply the knowledge we’ve learned in school very quickly.

And then the top challenge is… Medical literature is expanding at a staggering rate. Like you noted, the number of treatments has expanded at a staggering rate, but much more so, our ability to measure individuals has expanded. And as a result, even sort of knowing our notion of what is a disease…

It’s not just the case that… The rules aren’t so simple anymore. It’s much more challenging. Rather than saying, “For every person with sepsis, give them fluids.” No.

Some are very responsive, and some are not responsive, and the obvious one is if they have any kind of heart failure, don’t give them fluids because it’s going to make the condition worse. What I’m effectively going to is…

I feel there’s a huge low-hanging fruit here, which is… I think we can make human health a lot better by even thinking just harder about even all the treatments we already have, as we start taking many more measurements, and as these measurements are becoming visible to us in ways that they’re accessible.

Improving the precision at which we prescribe these measurements will make a huge difference, and I think that’s very tangible, very easy to… I think something we’ll get to within the next five to ten years. There are lots of areas of medicine that will see a huge improvement, just from better use of lots of data, that we already know how to collect. And thinking about the use of that data and improving how we target therapy.

I’ll give you an example: An area study that I am familiar with is, as I mentioned earlier, these complex diseases—like scleroderma.

They used to think of scleroderma as one disease, and any expert who treats scleroderma patients knows that there’s tremendous diversity among individuals when they come in. Some have huge impact on the kidneys, others have a huge impact on the gastrointestinal tract, and yet others have huge impact on the heart or lungs.

And effectively, when the persons come in, you’re kind of wondering, “Well, I have an array of medications I can give them. Who is this person going to be? And what should I be treating them with?” And our ability to look at this person’s detailed data and understand who this person is likely to be… And then, from that, targeting therapy more effectively, could already influence and improve treatment there.

So I think that’s one area where you’ll see a huge amount of benefits. The second area that I think… is basically increasing our ability to measure more precisely. And you can already see whole genome sequencing, microbiomes, and there are specific disease areas where being able to collect this much more easily will make a big difference.

And then, effectively, they’re going to give rise to new treatments because there are pathways that we are unaware of, that we will discover in the process of having these measurements, and that will lead to new treatment. So, I think the next ten years are going to be very, very exciting in terms of how quickly the field is going to improve. And human health is going to improve from our ability to administer medication and administer medicine more precisely.

That is a wonderful thought. Why don’t we close on that? This has been a fascinating hour and I want to thank you so much for taking the time to join us.

You’re welcome and thank you so much for having me! This was really fun.

Byron explores issues around artificial intelligence and conscious computers in his upcoming book The Fourth Age, to be published in April by Atria, an imprint of Simon & Schuster. Pre-order a copy here

.voice-in-ai-link-back-embed {
font-size: 1.4rem;
background: url(https://voicesinai.com/wp-content/uploads/2017/06/cropped-voices-background.jpg) black;
background-position: center;
background-size: cover;
color: white;
padding: 1rem 1.5rem;
font-weight: 200;
text-transform: uppercase;
margin-bottom: 1.5rem;
}

.voice-in-ai-link-back-embed:last-of-type {
margin-bottom: 0;
}

.voice-in-ai-link-back-embed .logo {
margin-top: .25rem;
display: block;
background: url(https://voicesinai.com/wp-content/uploads/2017/06/voices-in-ai-logo-light-768×264.png) center left no-repeat;
background-size: contain;
width: 100%;
padding-bottom: 30%;
text-indent: -9999rem;
margin-bottom: 1.5rem
}

@media (min-width: 960px) {
.voice-in-ai-link-back-embed .logo {
width: 262px;
height: 90px;
float: left;
margin-right: 1.5rem;
margin-bottom: 0;
padding-bottom: 0;
}
}

.voice-in-ai-link-back-embed a:link,
.voice-in-ai-link-back-embed a:visited {
color: #FF6B00;
}

.voice-in-ai-link-back a:hover {
color: #ff4f00;
}

.voice-in-ai-link-back-embed ul.go-alexa-briefing-subscribe-links {
margin-left: 0 !important;
margin-right: 0 !important;
margin-bottom: 0.25rem;
}

.voice-in-ai-link-back-embed ul.go-alexa-briefing-subscribe-links a:link,
.voice-in-ai-link-back-embed ul.go-alexa-briefing-subscribe-links a:visited {
background-color: rgba(255, 255, 255, 0.77);
}

.voice-in-ai-link-back-embed ul.go-alexa-briefing-subscribe-links a:hover {
background-color: rgba(255, 255, 255, 0.63);
}

.voice-in-ai-link-back-embed ul.go-alexa-briefing-subscribe-links .stitcher .stitcher-logo {
display: inline;
width: auto;
fill: currentColor;
height: 1em;
margin-bottom: -.15em;
}

Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s