Pascale Fung on building robots with empathy (15-04-2016)

Building robots and machines with empathy

More videos with Pascale Fung

automatically generated captions
00:00:00 Speaker 1: You think that empathy modules are indispensable for machines and robots of the future, why?
00:00:08 Speaker 2: Because we are going to have robots in our lives.
00:00:13 They are going to work with us, they are going to take care of us.
00:00:16 So for us to be able to trust them, we need to build a relationship with the robot.
00:00:22 Studies have show that for us to work with robots we do need robots to understand our intention and emotion
00:00:30 and not just what we say. It's not just a command and control kind of machine.
00:00:34 Speaker 1: Aren't current robots good enough in that respect?
00:00:37 Speaker 2: No, because they take your commands literally right now. They don't understand what you really mean.
00:00:45 You can say the same thing and in different tones and with different gesture it will mean something totally different.
00:00:51 Suppose you have a robot that is taking care of the elderly and the robot says "how are you doing today"
00:00:57 and you say "I'm not sure." If the robot understands you literally, it just means you are not sure
00:01:07 and the robot will just take off and walk away.
00:01:10 But it really means, what the [inaudible 00:01:12] really is that the patient
00:01:15 or the elderly person is not feeling that great.
00:01:18 So that robot has to deduce that intention and emotion from what you are saying,
00:01:23 how you are saying it in order to take care of us better..
00:01:26 Speaker 1: And currently robots aren't ..
00:01:27 Speaker 2: No they are not able to do that yet, we are working on making them to do that.
00:01:35 Speaker 1: What made you actually want to introduce empathy in robots?
00:01:39 Speaker 2: Because I believe that in the future we are going to need robots to take care of people,
00:01:48 to take care of the young, to educate us, to help us with a lot of work.
00:01:52 Since we're going to have them around us I think it is important that they're more human like in their empathy..
00:02:02 Speaker 1: Why are empathy modules indispensable for lets say .. Tomorrow's robots?
00:02:09 Speaker 2: We want robots to be intelligent, right?
00:02:14 Intelligent robots needs to have both the cognitive intelligence and also emotional intelligence.
00:02:21 This is what we humans have. When we communicate with each other, we use our emotional intelligence all the time.
00:02:30 That is indispensable for understanding each other.
00:02:33 For robots to understand us, they need to have that kind of emotional intelligence, which is empathy.
00:02:38 Speaker 1: Because robots aren't capable of doing that?
00:02:41 Speaker 2: No, not yet. Current robots is most of the time still controlled by explicit commands.
00:02:49 For example you can tell a robot to vacuum your room.
00:02:52 In some restaurants there are robot waiters that will bring you food and they are focused very narrowly on one task.
00:03:00 They are not that much more advanced than your vacuum cleaner now
00:03:05 or the rice cooker. Current robots don't have that kind of emotional intelligence but there are some robots.
00:03:11 They're putting this in some robots and we are starting to see them..
00:03:15 Speaker 1: Why would we need intelligent emotion .. Emotional intelligence for machines?
00:03:19 Speaker 2: So that they can work with us better. They can help us better.
00:03:25 If they need to take care of us, take care of our children, our elderly,
00:03:29 they really need to understand our true intent emotion in order to take care of us. Say if you go to the hospital
00:03:36 there's a nurse and what is a nurse does is not just to take your temperature and look at your vital signs
00:03:43 but also talk to you and see how you are doing and whether you need comforting, whether need water,
00:03:50 whether you need medicine at this point in time, so that requires emotional intelligence and it requires empathy.
00:03:56 Speaker 1: Is it because they are becoming closer to us or because there is a difference of environment?
00:04:02 Speaker 2: Yes the robots are coming more into our lives, into our daily lives and there will be more robots around us.
00:04:10 If they don't have emotional intelligence they are more likely to make mistakes and even hurt us.
00:04:16 Speaker 1: There are a lot of different kinds of emotions.
00:04:20 What kind of emotions should they express or should they at least recognize?
00:04:23 Speaker 2: For example the very first thing we're working on for robots to recognize includes whether the human is
00:04:31 happy, angry, sad, or frustrated or hesitating or even sense of humor.
00:04:40 One of my students is working on recognizing sense of humor.
00:04:44 Speaker 1: Let's talk about the range of emotion. Of course there are a different kinds of emotions.
00:04:50 What is the let's say the primary emotion that they should recognize?.
00:04:52 Speaker 2: The primary emotions are happiness, sad .. Sadness, and anger, and neutral.
00:05:00 So you have to be able to tell whether the person is happy. Happy means satisfied, not satisfied, not happy.
00:05:09 Sad, needs help, maybe frustrated.
00:05:14 Angry, so if the person is angry at the robot, the robot should have to do something in response..
00:05:21 Speaker 1: But people can be ironic or .. [crosstalk 00:05:25]
00:05:24 Speaker 2: Right, we're also working on for robots to understand a sense of humor and sarcasm,
00:05:32 we're working on that because we use humor and sarcasm in our daily communications to deflect the situation
00:05:42 or challenge or to make a conversation more friendly, to make things go more smoothly.
00:05:51 Robots need to learn to recognize that.
00:05:54 Speaker 1: Even for people it is sometimes difficult to recognize [crosstalk 00:05:58] a robot to recognize that..
00:05:58 Speaker 2: Indeed .. So we're teaching robots to watch a lot of comedy shows to learn sense of humor.
00:06:06 It's machine learning so we think that if we let the robot watch a lot of comedy shows
00:06:14 and observe how people communicate with each other, the so called big-data analytics and use machine learning,
00:06:21 then they will be able to learn..
00:06:23 Speaker 1: So you teach them actually by letting them watch ..
00:06:27 Speaker 2: We feed comedy shows and movies and yes. And people's daily communications.
00:06:36 A lot of YouTube videos, we feed these to the robot..
00:06:40 Speaker 1: Should it also express somehow ..
00:06:43 Speaker 2: Yes that's very, it is important.
00:06:47 Studies have shown that humans feel more related to a machine that has some facial expressions right?
00:06:53 That's why we don't really feel a connection to our refrigerator.
00:06:58 or our rice cooker because they are just machines. When you see a robot with a cute face, people start going ..
00:07:05 They start talking to the robot in a more human way. They go "so how are you doing" and stuff like that.
00:07:13 Embodiment of the machine with a robotic body with facial expressions is important.
00:07:18 That is also an important area,
00:07:21 different researchers are working on that to allow robots to generate appropriate facial expressions,
00:07:28 appropriate gestures, for example to say "hi" or to shake you hand and all that..
00:07:34 Speaker 1: Because of tone of voice ...
00:07:35 Speaker 2: [crosstalk 00:07:36] The tone of voice also from the [crosstalk 00:07:36] .. Is very important too
00:07:38 Speaker 1: Some of you say that a robot should not alone learn to understand the content of speak
00:07:44 but it should it also compare it with the way it is delivered. Can you explain it?
00:07:52 Speaker 2: So if I say something like,
00:07:54 "I'm really happy I'm going to work today," you probably think that I'm truly happy that I'm going to work today..
00:08:00 But if I say "I'm really happy I'm going to work today ..
00:08:06 " Even though the words I speak are the same, the emotion is totally different.
00:08:11 In the second case you know I'm being sarcastic and I'm not really feeling like going to work.
00:08:16 So the tone of voice is important. We use tone of voice in our way of expressing our meaning a lot.
00:08:23 Speaker 1: But how do you teach a robot to do that?
00:08:25 Speaker 2: So again the difference is we teach the robot to listen, again, to different kind of tone of voice.
00:08:34 We show a robot examples of the same sentences in angry voice, happy voice, and nervous voice, and frustrated voice.
00:08:43 We tell the robot "Look this is the kind of voice that expresses anger,
00:08:47 this is the kind of voice that expresses frustration" and the robot learns.
00:08:53 Speaker 1: Somehow we learned that as children.
00:08:55 Speaker 2: Yes we did, we do as children.
00:08:59 Speaker 1: How does the robot learn that?
00:08:59 Speaker 2: So children have multiple sensor input, right?
00:09:05 Children perceive the world not just by what people are saying, how they are saying, but they also look at things.
00:09:12 So right now we are working towards enabling robots to have multiple sensory input as well.
00:09:18 So the robot now learns from the speech you're saying, what you're saying,
00:09:25 but also learns from how you're saying it. We also enable the robot to look at your facial expression
00:09:30 when you say it so that enable the robot to learn better.
00:09:35 Speaker 1: As a child you learn by experience, a robot learns by experience?
00:09:40 Speaker 2: Yes.
00:09:44 Speaker 1: As a child you can be traumatized by experience.
00:09:46 Speaker 2: Right.
00:09:47 Speaker 1: What about robots? Is it possible somehow to get robot trauma?..
00:09:51 Speaker 2: Robot trauma I would say .. Robot can be mislead. Okay, if you feed robot some information ..
00:10:00 It's like how we teach our children. Robots can be traumatized too because they are machines.
00:10:06 Machine can have machine errors, either from the embedded coding, the program or from experience,
00:10:13 meaning the data they learned from.
00:10:15 Speaker 1: It's a different kind of trauma than the human trauma.
00:10:21 Speaker 2: They can be lead astray and then they can have errors. Is it different from human trauma?
00:10:29 It's different, machine trauma, I will say.
00:10:32 Speaker 1: Children can be dysfunctional because of they're trauma or can't behave very well.
00:10:39 I mean, what about robots?
00:10:41 Speaker 2: Robots can be dysfunctional, they can stop functioning properly if they have had a machine error.
00:10:47 If the error is caused by they're experience you can call that trauma. They can be traumatized into misbehaving indeed.
00:10:56 Speaker 1: It's going to be grumpy?
00:11:00 Speaker 2: Grumpiness means they work slowly or they don't respond friendly in the friendly manner,
00:11:09 they may not even answer. You can perceive that as grumpiness, then machines can be grumpy, yes.
00:11:15 Speaker 1: You also think that robots should apologize somehow?
00:11:19 Speaker 2: Of course!
00:11:24 Speaker 1: What do you mean by robots should be apologize?
00:11:26 Speaker 2: Robots make errors like humans do.
00:11:28 These days when there is a machine error, you might just see a blue screen and it just says,
00:11:34 "machine error 404." That is not very friendly, we cannot have robots do that.
00:11:39 Machines are bound to make errors, there's not going to a perfect machine that never makes errors.
00:11:45 Just like humans,
00:11:46 there are no perfect humans that never make any mistake. It is important for empathetic machine to apologize like
00:11:54 humans do because that will keep the communications smooth. That will keep the conversation continue with a human user..
00:12:01 Speaker 1: Because that is actually the most important thing to keep [inaudible 00:12:05]
00:12:06 Speaker 2: Indeed yes it is most important to keep the conversation smooth and the human/robot communication smooth.
00:12:11 Speaker 1: But somehow you suggest that you should feel happy with your robot. Is that correct, am I correct?
00:12:22 Speaker 2: Humans should feel happy
00:12:23 when they are communicating with the robot at least they can feel related to the robot in a friendly sort of way.
00:12:31 It is important for humans to feel that in order for us to trust the robot, to work with the robot.
00:12:37 Speaker 1: We develop a character if we are born we have sort of character or maybe genetically.
00:12:43 But we give out the character because, well experience we have.
00:12:48 Can a robot develop a character as well in your opinion, your model?
00:12:52 Speaker 2: So, when we built empathetic robots we are very careful in also designing the robot personality.
00:12:59 This is what we call robot personality. In the beginning we will be designing the personality.
00:13:06 This is similar to human predisposed personality we're born with. As we go on we also let the robot learn personality
00:13:16 from data, so from particular personality type.
00:13:21 For example,
00:13:21 a robot can imitate a particular persons personality type over time by observing how that person communicates with
00:13:28 other people. A robot can also be nurtured into developing it's personality.
00:13:35 Speaker 1: So you have a discussion and you use for example where is nature [crosstalk 00:13:39] nurture.
00:13:42 What about robot?
00:13:42 Speaker 2: Robot also. There's a nature and nurture. So nature, robot nature comes from our laboratories right?
00:13:49 Comes from the design of robot by humans, by engineers that's the nature..
00:13:55 That's when robot comes out of lab and robot has that personality. But then as we ...
00:14:02 Since robots have machine learning algorithms in them, they will also learn from particular kind of ..
00:14:08 You know, they will learn from the environment, and then they will continue to develop their personality. For example,
00:14:15 the beginning we ensure that the robot is not racist. That is predisposed what we design at the beginning.
00:14:24 Over time when the robot sees whatever is out in the world, there can be racist comments and all that
00:14:31 but the robot will reject that and will not absorb that into the robots personality.
00:14:35 The robot will make a judgment to say "oh, that is racist and I should not learn that.".
00:14:41 Speaker 1: That suggests that you put moral codes into the robot ..
00:14:44 Speaker 2: In the beginning, yeah, in the beginning we need to teach robot values and personality.
00:14:50 Speaker 1: Doesn't it depend on what kind of culture you sit in?
00:14:55 Speaker 2: Yes.
00:14:57 [crosstalk 00:14:57] So the people who work on robots, we all have this responsibility,
00:15:03 we're all like parents to robots.
00:15:05 Parents to our children, we indeed in the beginning we do teach our robots in certain ways, with certain code.
00:15:12 Then we let them run.
00:15:14 Very much like how we nurture our children when they reach adulthood we just let them go. In fact
00:15:20 when they're younger we send them to schools so we don't always teach them everything ourselves.
00:15:25 We send them to school, we send them out into the world, they learn from the environment,
00:15:29 so that's what we do with robots.
00:15:35 Speaker 1: You talk a little bit as if it's becoming a person somehow..
00:15:39 Speaker 2: Robot is .. So we're building robots to be a little bit more like us, more like a person indeed.
00:15:47 Because humans can communicate with another human being better. We cannot ask humans to speak the robot language right?
00:15:56 So we the robot to speak our language, and understand us.
00:16:00 With that the personality thing, the robot values,
00:16:05 that is all necessary in order for us to communicate with the robot better.
00:16:10 Speaker 1: Do robots have a sort of self image in terms of I can think about myself?
00:16:17 What do robots think about itself as well?
00:16:19 Speaker 2: The robot can certainly be taught to think about itself yes or shows the behavior of thinking about itself.
00:16:29 Meaning that robot can simulate the behavior somebody who's thinking about itself.
00:16:35 Whether their is consciousness within the robot we don't know because we don't understand consciousness.
00:16:40 Speaker 1: That's what everybody's saying, "yeah you can program it but it's always pre-programmed.".
00:16:45 Speaker 2: It's not always pre-programmed. There's certain things ..
00:16:48 So this is what I was trying to say, we pre-program our machine learning way.
00:16:53 So we pre-program Robots to learn, just like humans are also programmed to learn. Part of it is pre-programmed.
00:17:03 This is what we call nature.
00:17:06 That pre-programming also allows the robot to learn,
00:17:09 to pick up from the physical world from interacting with human beings for their knowledge
00:17:15 and for their personality even.
00:17:18 Speaker 1: How can we monitor that?
00:17:20 Speaker 2: Can we monitor them?..
00:17:24 Speaker 1: I mean we monitor people, we tell people to not behave like that .. Good impression ..
00:17:36 Sort of punishment on bad behavior for example..
00:17:38 Speaker 2: Different robots, depending on the purposes ...
00:17:40 If they are supposed to take care of patients and they make mistakes then we think there's going to be ..
00:17:46 There must be a machine error, then we will check the machine, we check the code
00:17:50 and try to spot that error. We don't really punish robot per se at this point
00:17:57 but it can be imagined that in some learning algorithms we can incorporate what is a reward
00:18:02 and what is punishment so that they learn proper things.
00:18:05 Speaker 1: For example there is a discussion about autonomous cars for example [crosstalk 00:18:10] who is responsible?.
00:18:13 When robots come closer you'll get the same kind of question because they can harm or somehow .. [crosstalk 00:18:20].
00:18:20 Speaker 2: Indeed, indeed ..
00:18:21 If they accidentally harm somebody you say "you know, robot you shouldn't do that" and they should learn from that.
00:18:29 Punishment perhaps not like human punishment, I don't we can hit the robot
00:18:33 and the robot will feel "ooh," not that kind of punishment.
00:18:37 It's in the algorithm that there's a cause function, whether they are doing the right thing or not.
00:18:43 There will be a cause function of positive and negative values.
00:18:46 Speaker 1: What's your wildest dreams in terms of robots?
00:18:48 Speaker 2: My wildest dreams is to have my memory, my sensory abilities, my intelligence,
00:19:01 my emotional intelligence whatever it is all be downloaded to a robot with an android body
00:19:09 and then that robot will continue functioning as me when I'm no longer in this world..
00:19:15 Speaker 1: You want to make a copy of yourself basically ..
00:19:18 Speaker 2: I want to make a copy of myself, that will be interesting.
00:19:21 Speaker 1: Are you think that's feasible?
00:19:23 Speaker 2: I think to some extent we're already doing that..
00:19:27 I'm not sure it's feasible within my lifetime but I think it is feasible. The robot will have ..
00:19:36 Will be equipped with our perception abilities, and our intelligence, our emotional intelligence,
00:19:41 and our ability to learn.
00:19:45 The people who build the robot with really like form with very lifelike skin and eyes and all that..
00:19:54 When we put this together the robot will have the ..
00:19:59 Will be embodied in an almost human like body so they will pick up signals from the world as we do.
00:20:06 I think that is feasible. I'm not saying that will be me, that will be just a copy of me
00:20:12 and that will not necessarily have the conscious of me. I'm not talking about me living forever.
00:20:17 I'm talking about a copy.
00:20:19 Speaker 1: But what would it be?
00:20:21 Speaker 2: What would it be, a very good question. A copy. Just enjoy as a copy of a human.
00:20:29 Speaker 1: What would a copy of you mean? What would it do?
00:20:35 Speaker 2: It would do what I would do under given circumstances.
00:20:41 For example, maybe it can go and lecture in the university, teach students.
00:20:45 It can learn like I do from the world, maybe it can perform research. It can build other robots..
00:20:52 Speaker 1: But it won't be you ..
00:20:54 Speaker 2: It will not be me it will be a copy of me.
00:21:00 Speaker 1: But suppose it would be a copy of you, would it develop it's own consciousness? It's quite hard to imagine.
00:21:08 Speaker 2: I know, I know. We don't know what consciousness is right. It's almost a philosophical question.
00:21:16 Does it exist?
00:21:19 Once we have all this kind of sensory input intelligence
00:21:23 and learning in place would there be a point where there is indeed a consciousness? I don't know. I don't know.
00:21:30 We don't know what consciousness it and where it comes from.
00:21:32 Speaker 1: In terms of the old philosopher, let's say that you could say that robots [inaudible 00:21:37]
00:21:36 I think so I am.
00:21:39 Speaker 2: Yes indeed. I think so I am. Indeed, indeed.
00:21:44 A robot can conclude that but even then the robot can have the behavior of someone with consciousness.
00:21:50 We still don't know what really has this "me-ness' this self consciousness.
00:21:56 Speaker 1: If you look, let's say in 25 years, what would robots life look like in your perception?.
00:22:06 Speaker 2: So in 25 years I believe there will be robots .. Not all robots, some robots will look just like us.
00:22:14 They will talk just like us and behave just like us. It's possible in 25 years.
00:22:19 They can move and gesture exactly like humans. Some of these robots, these are called androids will be very very um ....
00:22:28 You cannot tell the difference between them and humans. Then there are other robots.
00:22:33 We still need other robots to help us, other robotic machines like the vacuum cleaner.
00:22:37 The vacuum cleaner is not going to have a human head that would be kind of creepy.
00:22:41 So we still have other robots that don't look like humans, so there will be a variety of robots among us.
00:22:48 Speaker 1: You are working to make it into a reality.
00:22:52 When you expose this scientific work, what's the response from people?
00:22:57 Speaker 2: Some people feel threatened. Some people question why robots need to be more human like.
00:23:07 They think that then they will be challenging us and taking over us..
00:23:12 Others like especially children think it's just really cool. They want to .. They cannot wait for that to happen.
00:23:19 I think depending on the culture people come from and what they expect of machines, they have different reactions.
00:23:26 For the most part, for example I talk to doctors, medical doctors
00:23:30 and they love the idea that robots will have empathy towards patients so that they can better take care of patients.
00:23:37 Speaker 1: Can you imagine that people somehow become scared or afraid?
00:23:41 Speaker 2: Some people are already scared. Some people are already scared of robots.
00:23:47 I think when people see robots become more human like they imagine all sorts of things.
00:23:53 I think one reason is there has been too many science fiction movies that portray robots as threatening and menacing.
00:24:00 but that's just science fiction. People shouldn't be .. People should not be swayed by fiction.
00:24:09 Working under reality we're all building robots to help people.
00:24:13 Nobody's building robots to purposely destroy human kind and that kind of thing...
00:24:23 Speaker 1: [inaudible 00:24:26] .. Let's robots be compelling and empathetic ..
00:24:26 Speaker 2: We're trying to build robots that are friendly.
00:24:30 If robots have empathy, then they will never harm people, right? That's why empathy is important.
00:24:38 If they do have empathy they will never hurt us.
00:24:41 Speaker 1: There are some robots laws made by Isaac [inaudible 00:24:44] [crosstalk 00:24:46].
00:24:45 Speaker 2: The three laws of robotics .. I don't remember all of them..
00:24:52 Speaker 1: But actually one of them is to help people ...
00:24:56 Speaker 2: Yeah to help people and not to harm ..
00:24:58 Speaker 1: That's your purposes.
00:24:58 Speaker 2: Indeed, indeed. So one of the three laws of robotics is for robots to help people.
00:25:03 I think all the robots that are being worked on today are trying to do that.
00:25:08 Speaker 1: Could you imagine that people start to see robots, if they are so friendly as a friend? Really as a friend?
00:25:15 Speaker 2: I hope so, I hope people will see robots as friends because they're friendly.
00:25:20 If we can see robots as friends we can trust them to help us.
00:25:24 Speaker 1: They are friends forever? Companions somehow?
00:25:32 Speaker 2: We are building robots to be people's companion.
00:25:38 To be companions to children, to be companions to the elderly when they're lonely and all that.
00:25:46 Like all machines, the hardware deteriorates right?.
00:25:50 So, your iPhone or your smartphone, you might buy the next generation smartphone ..
00:25:55 So robots, are we going to have a robot that lasts forever? I don't think right now that's the purpose.
00:26:02 So you might have the next generation of the same robot,
00:26:04 but that robot will have the same personality as the first robot and will have the memory. So yeah,
00:26:10 in that sense you can have a robot companion forever. But the body might have to be changed from time to time.
00:26:18 Speaker 1: If you were wanting can it be copied to another robot or how can you imagine that?
00:26:25 Speaker 2: The intelligence of robot is completely software based. We can make multiple copies of the same software.
00:26:30 We can have the same robot, same personality, same memory, in different robotic bodies.
00:26:35 So we can have multiple robots that sound the same, behave the same, and do the same thing.
00:26:41 Speaker 1: And copy their experience?
00:26:44 Speaker 2: And copy the experience because it's software based.
00:26:47 Speaker 1: Somehow people love their machines like they love their cars.
00:26:53 They start loving their robot as well, probably.
00:26:56 Speaker 2: I think people will love their robots if the robots are friendly, empathetic, cool, have a sense of humor,
00:27:03 who wouldn't love them?
00:27:04 People will love them like they love their motorcycles
00:27:08 and their cars indeed. But they also might want to have multiple robots,
00:27:12 a different version with a different personality just like some people like multiple cars of different styles.
00:27:17 Speaker 1: Do you think we are prepared for those kinds of problems?
00:27:21 Speaker 2: I think people are prepared, we've already seen men on the moon 40-50 years ago.
00:27:29 It's high time we see robots among us and I think people are ready for robots as long as they are friendly.
00:27:34 Speaker 1: Language is very important and the intention of language. Explain a little bit of what you're doing..
00:27:43 Speaker 2: What we're doing is um .. For example us the example of humor as I explained before.
00:27:51 Humor comes from not just the words you use but the tone of voice and even your facial expressions.
00:27:58 The same word expressed in different context can be humorous and not humorous.
00:28:03 So what we do is,
00:28:05 we program machines to have learning algorithms so they learn from watching a lot of comedy shows for example
00:28:13 and YouTube videos
00:28:14 and figure out what humor is. So next time somebody says something humorous they machine will know
00:28:20 and to be able to laugh for example.
00:28:23 Speaker 1: How do they know that it's funny?
00:28:27 Speaker 2: How do they know it's funny? From learning.
00:28:28 The learning algorithms enable the machine to see examples, many many examples, millions of sentences..
00:28:37 Thousands and tens of thousands of TV shows where people laugh. For example, humor consists of ..
00:28:44 How do you tell a joke? Usually there's a setup, a trigger, and a punch line.
00:28:52 The machine will see that in all these comedy shows like humans do and then they will pick up
00:29:00 when there will be a punch line and they know that is humor. The machine algorithm,
00:29:04 learning algorithm we use currently is what is commonly known as deep learning algorithms.
00:29:10 Speaker 1: How does this robot learn what's humorous?
00:29:15 Speaker 2: We use machine learning to teach robots to learn about our emotion sentiments including humor.
00:29:22 Machine learning there are two kinds of machine learning approaches.
00:29:27 One is what we call supervised learning the other is unsupervised. Supervised learning we actually give machines
00:29:33 examples and we have the data annotated by humans.
00:29:38 Humans say, "look, this is a sentence that's humorous," "this is a sentence that's a punchline for a joke,"
00:29:43 and "this is where people laugh." That's called supervised learning
00:29:46 and machines learn from that to have a sense of humor. Unsupervised learning is more like how humans learn.
00:29:53 We don't tell the machine explicitly this is a humor, this is not humorous, we give the machine a lot of data
00:30:00 and for the machine to learn from context.
00:30:03 So the unsupervised learning is really what we hope to achieve in the future.
00:30:07 If machine can have unsupervised learning, then we don't need humans explicitly teaching machines all the time,
00:30:14 this is humor, this is happiness and all that. That would save a lot of effort from the human annotators.
00:30:19 Unsupervised learning is harder tho because it will require a lot of general learning abilities
00:30:27 and general deduction abilities, induction. I believe it will require machines to have multi-sensory input.
00:30:36 Speaker 1: Do we know how unsupervised learning is working?.
00:30:39 Speaker 2: So cognitive scientists have been studying how humans learn. They believe that humans have ..
00:30:48 Of course we are born with some kind of learning abilities, innate.
00:30:52 When babies are born they already know how to recognize their mothers faces and voices.
00:30:59 They are already picking up from 10 months in the mothers tummy..
00:31:03 Speaker 1: Lets just [inaudible 00:31:05] .. Do we know how people learn unsupervised?
00:31:09 Speaker 2: We don't know exactly how we learn unsupervised. We're trying to model that.
00:31:15 So another side of working on robots is that
00:31:18 when we want to build robots to be more human like we have to build models of human thinking.
00:31:24 As we research how to make robots that have a sense of humor we understand human sense of humor better.
00:31:32 As we learn, as the machines learn, we also learn how we function.
00:31:38 We don't know exactly how we learn in unsupervised fashion but we're trying to research on that.
00:31:47 That's a very important research direction.
00:31:50 Speaker 1: There's a lot of talk about artificial intelligence,
00:31:55 what is let's say the condition of having an artificial intelligence system in your opinion...
00:32:02 Speaker 2: The condition of artificial intelligence .. We are trying to ..
00:32:06 Working towards what some people call a strong AI which is a general purpose AI system or general purpose robot.
00:32:14 Today we're only working, we're still only working on single purpose or multipurpose robots that can do one task.
00:32:22 You've heard of [inaudible 00:32:25] beating world champion [inaudible 00:32:27]
00:32:28 and their systems that can lift heavy weights and assemble cars. These are all single purposes robotic systems.
00:32:36 So we are working towards a general purpose robotics system that can really be your companion and take care of people.
00:32:42 In that case, the robot must have intelligence to be more human like. For human intelligence, empathy is indispensable.
00:32:51 It must recognize not just what people are saying but how people are saying it, what you truly mean.
00:32:59 So empathy is important. Emotional intelligence is a huge part of human intelligence.
00:33:04 Speaker 1: You suggest then couldn't be intelligence without [crosstalk 00:33:08]
00:33:07 Speaker 2: We cannot call a system or robot, human like intelligence without empathy.
00:33:14 Speaker 1: Is that what is lacking somehow in the discussion?
00:33:19 Speaker 2: It has been lacking in our discussion of building AI systems for the last 30 years, but it's coming now.
00:33:29 I'm happy to say I've been talking about it and some other researchers are also working on this..
00:33:36 Speaker 1: And because you are working in this field [crosstalk 00:33:38]
00:33:38 Speaker 2: That is a new direction and people in general agree with this direction so yeah,
00:33:46 research is working towards this goal as well. There is agreement in terms of that.
00:33:51 Speaker 1: Can you imagine people being afraid of you know, this huge intelligent network that is going to conquer us
00:33:58 or something like that?
00:33:59 Speaker 2: I think people are afraid of the unknown. People are always afraid of the unknown.
00:34:05 I think if we go back in time, 1950's and describe today's internet and smartphones and how we use them
00:34:14 and how we can get access to good material as well as bad material at your fingertips.
00:34:19 If we told people in the 50's about this today,
00:34:23 they would also be very very afraid. What happens is that we adapt to technology just like technology adapts to us.
00:34:31 It doesn't happen overnight.
00:34:34 We've been living with artificial intelligence for long time already starting with all those automatic calculators
00:34:42 and then we take airplane without being afraid that it's being flown a computer actually. So we've been living with
00:34:51 artificial intelligence for a long time its just gradually we are going to get used to the evolution of such
00:34:58 intelligent machines.
00:34:59 They will gradually be able to talk to us and they will gradually be able to empathize with our feelings
00:35:05 and they will gradually be able to do more to help us.
00:35:09 We'll get used to them step by step, it doesn't happen overnight.?
00:35:12 Speaker 1: What would be the ultimate condition to generate, let's say artificial intelligence [inaudible 00:35:26]
00:35:29 We already [inaudible 00:35:28] it's no problem.
00:35:30 Can you talk about robots with a heart, can you talk a little more about the necessity of this heart?
00:35:35 Speaker 2: Yes, so without empathy in the robot a robot would never behave and learn and understand like a human being.
00:35:45 So without empathy I would say that the robot would not be human like and that intelligence would be limited.
00:35:53 Speaker 1: So you think it is of the utmost importance of the entire discussion about artificial intelligence also.
00:35:56 should include ..
00:35:57 Speaker 2: The entire discussion of artificial intelligence does include that today.
00:36:07 I'm one of the people who champion it
00:36:11 and in general there is a general agreement that it is needed for artificial intelligence.
00:36:17 People work on different components of artificial intelligence.
00:36:21 Those of us work on the emotion recognition certainly see this as our job to make that happen,
00:36:30 to make robots have empathy.
00:36:32 Speaker 1: If you could lets say, look further
00:36:35 and further into the future can you imagine there is a combination of humans and robots
00:36:45 and artificial intelligence that goes beyond the stars for example or whatsoever?
00:36:48 It's larger than we are [inaudible 00:36:55] it's intention is [crosstalk 00:36:57] can you imagine?
00:36:53 Speaker 2: Yes what's interesting today, what's happening already is my mind
00:37:05 and your mind are no longer limited by our own life experience.
00:37:11 Twenty years ago I wouldn't know how to respond to a lot of questions that I'm not an expert in.
00:37:17 Today anybody with access to the internet
00:37:19 and Wikipedia can tell you a lot of things about a specific topic. Our human mind and our human knowledge has evolved.
00:37:28 We're already connected to this vast network of minds so you can pull up YouTube video to learn how to cook any kind of
00:37:38 food.
00:37:39 You can pull up Wikipedia page to learn about any particular technical topic or political topic or some history
00:37:48 and that happens instantaneously. So is that part of my mind already or part of the world? We already are connected.
00:37:56 So in the future when robots enhance our physical abilities they will also enhance our mental abilities.
00:38:03 When that happens, there will be basically, on top of the internet we have access to,
00:38:08 we also have these robots that enhance our ability to understand the knowledge.
00:38:15 That will be another layer of intelligence that enhances human intelligence.
00:38:20 Just like today there's robotic systems that help people who cannot walk, walk.
00:38:27 Those exoskeleton robots that can help people become stronger physically.
00:38:34 Robots can also enhance our intelligence to enable us to know more to be able to do more and think better,
00:38:41 with the the help of the robots.
00:38:45 Speaker 1: Earlier on I asked you your wildest dreams can you explain a little bit more intensely what your wildest
00:38:52 dream is?
00:38:53 Speaker 2: To build robots with all the components of human intelligence, human learning abilities, human perception,
00:39:04 human memories and human judgment, human values. So a long list of these things.
00:39:14 My wildest dream will be able to do that and teach that to robot. For example for a robot to copy all that from me.
00:39:22 My personal experience, my memory, my judgment, my values, which evolve as well. It's like a copy of a person, of me.
00:39:35 When I'm not around that copy will continue, maybe continue to talk to my children's children.
00:39:42 They know it's not me, they know it's not mommy but its a copy.
00:39:45 That will be interesting to replicate before I die so I see if it's really me or not..
00:39:53 Speaker 1: It's a replica? That's actually what you ..
00:39:55 Speaker 2: A replicant? Yeah you've seen Blade Runner right? I'm a huge fan of Blade Runner.
00:40:03 Yeah my wildest dream will have replicants but replicants also know they are replicants.
00:40:08 They don't fool people into thinking they are human beings.
00:40:16 Speaker 1: [inaudible 00:40:20]
00:40:17 Speaker 2: It doesn't have to be me, it can be anybody.
00:40:22 Speaker 1: If you copy let's say your intentions, your information somehow, what would it be? Would it be you?
00:40:32 What do you think?
00:40:34 Speaker 2: I don't know, I'm curious to know. If that happens would that be me?
00:40:41 Would that be just a copy of me? I can say that today we have people who can build a robot that physically looks like
00:40:47 me, exactly a copy of me but intelligent wise, memory and all that its not close.
00:40:55 It's still very very far from being a complete copy of a real human being. If we have more of a almost perfect copy
00:41:05 would that still be just a copy? I think its just a copy, its not me.
00:41:10 Speaker 1: It's not a wish for immortality?
00:41:12 Speaker 2: No, it would be my avatar I would say.
00:41:13 That would be avatar at the physical and mental level but still an avatar.
00:41:22 Speaker 1: Not the real you.
00:41:24 Speaker 2: Not the real me. Maybe it can do the tasks that I can do, I don't know.
00:41:29 Maybe it can continue to teach but it would be an avatar.
00:41:33 Speaker 1: Would it be something like a 3D or 4D photograph of you somehow for your children or grandchildren?
00:41:42 Speaker 2: I am actually thinking it's going to be a physical body with very human like skin
00:41:48 and very human like everything.
00:41:51 There are people working on building that so I think it's entirely possible in 20 years..
00:41:58 The body is possible, but the mind, because we don't understand the mind completely so how ..
00:42:06 Component by component,
00:42:07 module by module we're building the mind into the robot Today were still talking about service robots
00:42:15 and home robots that do particular tasks. We're still not building a general purpose human like robot like a replicant.
00:42:23 We're not doing that yet.
00:42:25 We don't have an interest in doing that it would be more like a scientific pursuit because we don't know what we need.
00:42:33 Why do we need a general purpose robot, that is exactly like me? What is the application of that?.
00:42:39 There's no application except that ..
00:42:40 That would serve us like a scientific quest rather than engineering application, rather than commercial purpose,
00:42:49 its a scientific quest.
00:42:51 As I mentioned earlier, when we study robot,
00:42:53 when we model intelligence for robot we are also modeling human intelligence, we're studying humans at the same time
00:43:01 and that is interesting to me.
00:43:02 Speaker 1: Some people could say she has been reading too many science fiction.
00:43:08 Speaker 2: (laughter) yes some people will say that.
00:43:09 But you know, many people who work artificial intelligence are huge science fiction fans.
00:43:14 We are just naturally attracted to science fiction since we were young
00:43:18 and then we got into this area because of that. Many of us are, many.
00:43:22 Speaker 1: To make it into reality?
00:43:24 Speaker 2: Yeah, to make it into reality.
00:43:26 It's fair to say that a lot of our imagination are shaped by science fiction we grew up with.
00:43:34 So you will see things that look science fiction like and its not a coincidence it's just we're shaped by that.
TV Get inspired and watch tv episodes of The Mind of the Universe, made by Dutch public broadcaster VPRO
  • Browse through over 30 hours of interviews
  • Download the interviews, including subtitles
  • Remix, re-use and edit under CC-BY-SA license
  • Start exploring