1 00:00:01,280 --> 00:00:05,570 Hi, everybody. My name is Mike Wooldridge and I'm a professor of computer science at the University of 2 00:00:05,570 --> 00:00:09,830 Oxford and I'm currently head of Department of Computer Science at the University of Oxford. 3 00:00:09,830 --> 00:00:14,840 And it's my very great pleasure to welcome you to our Hillary term stretchy lecture. 4 00:00:14,840 --> 00:00:22,640 The second straight lecture that we've done online, we hope it will be only the second or third and that by November we will, 5 00:00:22,640 --> 00:00:26,390 with luck, be back to being able to do them face to face. 6 00:00:26,390 --> 00:00:38,030 So before I begin, I would just like to I would just like to thank our sponsors and our sponsors are Oxford Asset Management. 7 00:00:38,030 --> 00:00:44,900 They've been supporting these lectures now for six years and they've completely transformed what we've been able to do with these lectures. 8 00:00:44,900 --> 00:00:49,700 Obviously, we've been in a lockdown now for the past year. But when we went into lockdown, 9 00:00:49,700 --> 00:00:53,690 they enabled us to bring speakers that we couldn't have imagined bringing from 10 00:00:53,690 --> 00:00:57,860 across the world and completely transformed our structured lecture series. 11 00:00:57,860 --> 00:01:02,360 So we are tremendously grateful to you all at Oxford Asset Management. 12 00:01:02,360 --> 00:01:08,990 I hope you will be. I hope you will be tuning in and joining us today and and for the future, I should say. 13 00:01:08,990 --> 00:01:14,240 They always ask me to say this. They are hiring. They are based in Oxford. If you're interested in machine learning. 14 00:01:14,240 --> 00:01:20,030 If you're interested in advanced computational techniques in Oxford, Asset Management is definitely worth checking out. 15 00:01:20,030 --> 00:01:28,670 And I'd given you the you are out there. So onto today's lecture. 16 00:01:28,670 --> 00:01:32,420 So the hour straight you lecture today is Subbarao come home Bartee. 17 00:01:32,420 --> 00:01:35,630 He is a professor of computer science. Arizona State University. 18 00:01:35,630 --> 00:01:42,920 And he studies fundamental problems in planning and decision making and is motivated particularly by the challenges of human aware A.I. systems. 19 00:01:42,920 --> 00:01:48,860 He holds a number of prestigious fellowships, including a fellowship from the Association for the Advancement of Artificial Intelligence, 20 00:01:48,860 --> 00:01:54,230 triple-A AI, the American Association for the Advancement of Science, the Association for Computing Machinery. 21 00:01:54,230 --> 00:02:03,020 And he was a NSF young investigator. He served as president of triple-A on the Association for Advancement of A.I. and has served as a trustee 22 00:02:03,020 --> 00:02:08,630 of the International Conference on A.I. and has been a founding board member of the Partnership for A.I. 23 00:02:08,630 --> 00:02:17,450 His views on research, as well as progress and societal impact on A.I. are featured in many multiple international media outlets. 24 00:02:17,450 --> 00:02:22,760 And in particular, he writes a column, a regular column full on artificial intelligence for the Hill, 25 00:02:22,760 --> 00:02:26,240 which is both very entertaining and extremely informative. 26 00:02:26,240 --> 00:02:38,240 And I recommend it if you're interested in learning more about the reality of A.I. today and in seeing some commentary on any AI developments. 27 00:02:38,240 --> 00:02:45,140 He's also on Twitter and is an extremely entertaining person to follow on Twitter, 28 00:02:45,140 --> 00:02:51,800 making lots of interesting remarks about current developments in A.I. so raw. 29 00:02:51,800 --> 00:02:53,600 It's all very great. Pleasure to welcome you. 30 00:02:53,600 --> 00:03:00,650 I'm sorry we can't welcome you face to face, but as I just said to you off line, we hope that we will be able to do so at some point in the future. 31 00:03:00,650 --> 00:03:05,330 And you would be most welcome. So, Ralph, over to you. 32 00:03:05,330 --> 00:03:17,660 Thank you. So thank you all for coming for this lecture from wherever you are. 33 00:03:17,660 --> 00:03:25,220 And thanks, Mike. And Oxford as this department for inviting me to give this lectured. 34 00:03:25,220 --> 00:03:37,830 So. I am going to be talking to you today about how to get a chance to interact and collaborate with us on our terms. 35 00:03:37,830 --> 00:03:51,390 So let me get started by thinking about how we basically CEI being shown being represented in popular culture. 36 00:03:51,390 --> 00:03:57,480 Typically, AI systems already from the beginning from the get go. 37 00:03:57,480 --> 00:04:07,140 I see it as things that are like companions to humans. Whether you start from like 2001, A Space Odyssey, a HAL, you know, hanging around with Dave, 38 00:04:07,140 --> 00:04:14,370 Samantha hanging around in HUD and also, of course, our in ex machina even. 39 00:04:14,370 --> 00:04:23,620 Obviously, the popular culture, the novels, you know, Ian McKellen and most recently Cosla should go to talking about robots that work, 40 00:04:23,620 --> 00:04:28,260 you know, the way they imagine the baby imagining agencies. They will be working with us. 41 00:04:28,260 --> 00:04:35,820 They'll be sort of companions. And, you know, you will almost treat them almost as if that humans at some level. 42 00:04:35,820 --> 00:04:41,800 So this is the I in popular culture. So you would think that, you know, instead of the gloom and doom about it, 43 00:04:41,800 --> 00:04:47,790 I see the areas you'd be seeing this kind of breaking news in popular culture. 44 00:04:47,790 --> 00:04:51,720 Technically, I that is helping old lady across the street. 45 00:04:51,720 --> 00:04:55,920 It plays with kids, cooks, food hacks at home and sense of drama. 46 00:04:55,920 --> 00:05:01,880 In case you're wondering, how come we never saw this kind of breaking news item on CNN? 47 00:05:01,880 --> 00:05:04,920 It's because I made it up. It's fake news. 48 00:05:04,920 --> 00:05:14,010 Basically, it turns out that we typically keep hearing about a I not in the sense of agents working with humans. 49 00:05:14,010 --> 00:05:23,220 So there's been essentially a sort of a curious ambivalence that my research has had with humans. 50 00:05:23,220 --> 00:05:32,670 Our system seems to be happy as to either far away from humans, like the spirit, opportunity, curiosity. 51 00:05:32,670 --> 00:05:41,400 I'd basically in some adversarial stems with humans essentially claiming the heck out of us in various kinds of board games. 52 00:05:41,400 --> 00:05:46,260 It's almost as if we wanted to help humanity. 53 00:05:46,260 --> 00:05:54,150 It's the people that we idee, such as just gonna stand paraphrasing something that John Lennon said. 54 00:05:54,150 --> 00:06:00,900 Right. So I was sort of. This is something. Been something that has been weighing on me for several years. 55 00:06:00,900 --> 00:06:12,540 And I've seen sort of my research programme in for the last several years has been in terms of what the challenges of human activity ISIS stems. 56 00:06:12,540 --> 00:06:19,560 How do you get human systems to come? You know, I collaborate with us as companions. 57 00:06:19,560 --> 00:06:28,190 So I give a talk at the U.S. back at Tripoli Eye that sort of set up as a magazine article on the challenges of Human A very I. 58 00:06:28,190 --> 00:06:33,360 And most recently, that whole thing that Mike mentioned, 59 00:06:33,360 --> 00:06:42,540 I wrote a column sort of giving the current state of the area in the I fight like Lape Oblique. 60 00:06:42,540 --> 00:06:47,030 What I want to do in today's talk is sort of give you some of the technical background. 61 00:06:47,030 --> 00:06:54,690 And as much as my group has been working on in some of making some of these challenges come to fruition. 62 00:06:54,690 --> 00:07:01,590 So I was sort of asking in the in my trip, I talk back in 2000. 63 00:07:01,590 --> 00:07:07,140 Why isn't human having any AI all over the place already, given that pretty much everybody thinks about the, 64 00:07:07,140 --> 00:07:14,310 I think, SBI agents as working with humans, at least outside of it, and why we should be pursuing it. 65 00:07:14,310 --> 00:07:21,580 And of course, you know, some of the challenges and one of the points I kind of made there, 66 00:07:21,580 --> 00:07:29,100 and that's what sort of repeating here is having for high research to take humans into account. 67 00:07:29,100 --> 00:07:33,180 So we have our agents work with humans in the loop. 68 00:07:33,180 --> 00:07:37,320 It's not sort of a dilution of our goals. 69 00:07:37,320 --> 00:07:44,430 It's really sort of an expansion of our goals of the scope and reach of the enterprise. 70 00:07:44,430 --> 00:07:52,230 After all, it's been said that we had developed evolutionarily these crazy large breeds. 71 00:07:52,230 --> 00:07:58,450 We have not to run away from the lands of the Savannah. 72 00:07:58,450 --> 00:08:05,420 Are the title staff loyal Bengal. But to really strategize deal with each other. 73 00:08:05,420 --> 00:08:11,400 Also, this social interaction is what really potently led to some of the brain size. 74 00:08:11,400 --> 00:08:16,800 And please, at least speculatively, often from an evolutionary theory. If you don't believe that, of course, 75 00:08:16,800 --> 00:08:21,940 we do know that we that we are always modelling each other's mental states and what 76 00:08:21,940 --> 00:08:26,420 the other people are thinking about me and what would I do to counteract that? 77 00:08:26,420 --> 00:08:31,780 There's a great piece from like French, the show that some of you may have, 78 00:08:31,780 --> 00:08:38,280 Watchmen, Phoebe, saying they don't know that, we know that they know that vino. 79 00:08:38,280 --> 00:08:48,320 So you can actually play with these mental models both to cooperate and to strategize against, you know, adversarial stances. 80 00:08:48,320 --> 00:08:55,140 I mean, obviously, human eye systems are like needed in many places from the very beginning. 81 00:08:55,140 --> 00:09:01,490 Actually, people didn't realise that things like intelligent tutoring systems on the social robotics, 82 00:09:01,490 --> 00:09:06,210 the researchers in those areas have certainly taken humans into consideration 83 00:09:06,210 --> 00:09:10,200 in designing their systems because that's the whole point of those systems. 84 00:09:10,200 --> 00:09:18,090 But then that's not the only pieces where you have you want your as systems to work with humans, 85 00:09:18,090 --> 00:09:22,800 pretty much even quotidian interactions between systems for humans, 86 00:09:22,800 --> 00:09:30,810 for example, assistance, human digital, personal assistants, office hospital assistants and AFCO in teaming. 87 00:09:30,810 --> 00:09:40,760 You know, some of that is possible, Miles, some probably more so in the near future, elbow to elbow teaming more cognitive deeming, 88 00:09:40,760 --> 00:09:51,840 such as search and rescue scenarios and even like factory robotics, which can actually explain their operation to new humans in the loop. 89 00:09:51,840 --> 00:10:02,370 OK. So in some sense, increasingly, if in fact, that you believe the hype that yeye is everything and I actually each computer science, 90 00:10:02,370 --> 00:10:07,260 then human computer interaction really will become human interaction. 91 00:10:07,260 --> 00:10:14,370 And so we really should understand some of the challenges there and how to solve those in 92 00:10:14,370 --> 00:10:20,580 within the United States that I'm pretty familiar with the U.S. national priorities for. 93 00:10:20,580 --> 00:10:25,950 I include like 10 areas. And number two is your money AI interaction. 94 00:10:25,950 --> 00:10:33,990 And then there, like several research programmes, for example, Darbar has assessed ADAPT and a more recent one, BGT and so on. 95 00:10:33,990 --> 00:10:40,770 Many other programmes looking at these kinds of challenges. So getting to this stock, what do I want to say? 96 00:10:40,770 --> 00:10:53,500 What does it. What I really want to get across is what does it take for any AI agent to show interpretable behaviour in the prisms of human agents? 97 00:10:53,500 --> 00:10:57,540 And the simple answer that it turns out is managing mental model. 98 00:10:57,540 --> 00:11:03,390 Some of you have said, obviously, that's kind of obvious, but some of you may not have thought about this. 99 00:11:03,390 --> 00:11:11,120 This is something that psychologists have always been talking about. You know, something like Sallyanne past that some of you may be familiar with. 100 00:11:11,120 --> 00:11:21,030 It points out that kids recognise that what they believe about the world is different from what others believe about the world. 101 00:11:21,030 --> 00:11:24,050 And this is when they realise that they actually can lie. 102 00:11:24,050 --> 00:11:29,130 You know, not only they can cooperate by actually making clear what they believe in, but also lie. 103 00:11:29,130 --> 00:11:39,510 And, you know, I joke around that. I was extremely happy when my son started telling his fossilise like 25 years back because 104 00:11:39,510 --> 00:11:45,810 I knew he had a working brain which essentially is able to model other people's mind, 105 00:11:45,810 --> 00:11:50,370 mental states and realise that what he believes in is different from what others believe in. 106 00:11:50,370 --> 00:11:54,900 So managing mental models becomes that important aspect. 107 00:11:54,900 --> 00:11:58,230 So that gets me to the overview of the talk, 108 00:11:58,230 --> 00:12:06,510 which is sort of a perspective on how to do this from a bunch of research that we have been doing in the past. 109 00:12:06,510 --> 00:12:14,730 You know, six or seven years. So so the all you have to talk is I basically want to get you to see that effective 110 00:12:14,730 --> 00:12:19,920 human interaction requires systems to be able to manage human mental models. 111 00:12:19,920 --> 00:12:28,380 These include include not only the humans model of their task, but also humans model of the agents of that robot. 112 00:12:28,380 --> 00:12:32,880 So this is at least a second level. In general, mental models can have infinite regress. 113 00:12:32,880 --> 00:12:38,820 But psychologists have basically talked about the fact that humans tend to use about three level nesting. 114 00:12:38,820 --> 00:12:48,450 And I will specifically focus today on the second level of nesting because that winds up leading to explicable behaviour as well as explanations. 115 00:12:48,450 --> 00:12:54,390 I'm going to get to that in a minute. And then managing mental models brings up both influence and learning challenges. 116 00:12:54,390 --> 00:13:00,510 I'll talk about how we deal with them and then frameworks for human interaction, 117 00:13:00,510 --> 00:13:05,460 because we are a bunch of computer scientists and we might think we are humans. We know what humans want. 118 00:13:05,460 --> 00:13:13,500 And so here is a solution that is not going to be up to snuff, because essentially, you know, 119 00:13:13,500 --> 00:13:21,570 you really need to have a systematic race of evaluating whether your activities offer assistance and so on. 120 00:13:21,570 --> 00:13:26,560 Are actually that acceptable to the humans through systematic. 121 00:13:26,560 --> 00:13:36,970 Subject studies, and then finally, I'll end it with by pointing out that mental modelling capabilities allow agents not only to cooperate, 122 00:13:36,970 --> 00:13:43,450 just like they allow kids not only to cooperate with the humans, but also to manipulate the humans in the loop. 123 00:13:43,450 --> 00:13:46,180 In fact, people are worried about deep six. 124 00:13:46,180 --> 00:13:53,500 Eventually, we have to start worrying about head fakes, which is the agent making you believe something that's not actually true. 125 00:13:53,500 --> 00:13:59,260 And using that to an advantage that's kind of not beneficial to you. 126 00:13:59,260 --> 00:14:04,250 Again, I'm not saying that you're going to be out there trying to design any agent that are. 127 00:14:04,250 --> 00:14:06,520 But they can be hijacked by somebody else. In fact, 128 00:14:06,520 --> 00:14:14,470 I would argue that huge amounts of problems with Facebook and other social media are caused by the fact that people 129 00:14:14,470 --> 00:14:23,020 didn't quite realise that they are talking to something like something and an agent that is actually profiling them. 130 00:14:23,020 --> 00:14:30,280 When we talk to other humans spent already on guard, people who are putting their hearts out to these Facebook, giving away lots of information. 131 00:14:30,280 --> 00:14:35,830 And then, of course, I got used misused in multiple ways as we found out. 132 00:14:35,830 --> 00:14:41,500 So it introduces several new ethical challenges that we need to be worried about. 133 00:14:41,500 --> 00:14:51,120 Before I go into details, let me show the people who actually do most of the work that I'm talking about, which is my group of students right here. 134 00:14:51,120 --> 00:14:55,900 And so most of the work that I'm talking about has done by one of them, 135 00:14:55,900 --> 00:15:02,420 in particular on a car who just graduated this different year at this is this week and 136 00:15:02,420 --> 00:15:09,250 shuddered and thought that that fellow who got in I tend to watch award this year. 137 00:15:09,250 --> 00:15:14,200 He graduated a couple of years back that they're actually done a lot of this work that I'll be discussing. 138 00:15:14,200 --> 00:15:23,830 Other people also. OK. So let me start with how we handle this sort of human interaction. 139 00:15:23,830 --> 00:15:27,880 And it starts with, as I say, it's about mental models. So let's start with three. 140 00:15:27,880 --> 00:15:35,920 A tale of three models. When I say I know for the rest of the talk, when I talk about models, I just want you to understand that. 141 00:15:35,920 --> 00:15:38,680 I'm not wedded to the specific language, 142 00:15:38,680 --> 00:15:47,500 although a lot of our work was done with relational representations in PDL are planning domain description language. 143 00:15:47,500 --> 00:15:55,030 They've also done work with other kinds of models, such as dynamic programming models have been used in MGP are not all communities. 144 00:15:55,030 --> 00:15:59,680 These are all connected. Anyway, the important thing is what's the content of these models? 145 00:15:59,680 --> 00:16:05,110 You would think you can think of the contents of these models essentially as including InĂ¡cio state of the world. 146 00:16:05,110 --> 00:16:10,510 The goal state the actions, the observation capabilities, as well as the current plan. 147 00:16:10,510 --> 00:16:17,350 If any of either of the agents and we have two agents that they're talking about, the robot and the human asset, said robot. 148 00:16:17,350 --> 00:16:21,190 But that doesn't. I don't particularly care about embodied A.I. systems alone. 149 00:16:21,190 --> 00:16:24,900 I will actually show you that we work also with us. 150 00:16:24,900 --> 00:16:30,580 Basically, our software AI systems, software artefacts which are supporting us may not be embedded. 151 00:16:30,580 --> 00:16:38,230 So I'll just use the word robot. And in what general sense? OK. So that's what you have to keep in mind when we're talking about models. 152 00:16:38,230 --> 00:16:44,050 This can be presented, obviously, PDA. But they can also be represented in your favourite languages. 153 00:16:44,050 --> 00:16:50,770 So let's start talking about, like, the simplest single agent which is living in the world. 154 00:16:50,770 --> 00:16:58,940 And then as the agent is actually just by itself, it really has some approximate model, 155 00:16:58,940 --> 00:17:02,890 an executable model of the world, what it hopes is the executable model of the world. 156 00:17:02,890 --> 00:17:09,340 OK. We'll call that and you can use that to make its own course of action and live in that world. 157 00:17:09,340 --> 00:17:13,870 So you basically given its reward metrics, given its goals. 158 00:17:13,870 --> 00:17:17,080 It can actually come up with courses of action that are consistent. 159 00:17:17,080 --> 00:17:22,240 I didn't do it on a model and then hope that they can be executed in the world and start executing them. 160 00:17:22,240 --> 00:17:30,820 So this is something that we know in a one on one. The moment you put a human in the loop, which is another agent. 161 00:17:30,820 --> 00:17:36,490 But it's a not another artificial agent. Is this human agent with all the human fables. 162 00:17:36,490 --> 00:17:44,380 So one of the first things that would happen is that the human race ends up having a model, MFH, for the task. 163 00:17:44,380 --> 00:17:52,510 That means they have their own goals, their intentions and their observational constraints and so on. 164 00:17:52,510 --> 00:17:59,620 And so the robot to be able to at least sort of it, you know, at least not get they may not get into there. 165 00:17:59,620 --> 00:18:06,750 We are hopefully try to help them. It needs to have some idea of this image and they'll call that image. 166 00:18:06,750 --> 00:18:13,810 OK. So this image is a robot of the eye injuries approximation of the humans model. 167 00:18:13,810 --> 00:18:20,290 That's the first of the mental models we're talking about. And you heard about this in many places. 168 00:18:20,290 --> 00:18:26,740 When you talk about basically robots, are agents anticipating. 169 00:18:26,740 --> 00:18:34,000 Human behaviour. This is what we mean. We are anticipating human behaviour by using MHR. 170 00:18:34,000 --> 00:18:38,200 This can lead to assistance that keeps getting out of the way. 171 00:18:38,200 --> 00:18:41,980 So why do the humans so that you don't get out of it. Don't get into that. 172 00:18:41,980 --> 00:18:50,290 And sometimes also to support Oeming, et cetera. So this has been actually done quite a bit in the community. 173 00:18:50,290 --> 00:18:57,640 In fact, I was looking at the strategy lectures, and I know that I think a couple of years back I started Glowers, 174 00:18:57,640 --> 00:19:03,850 gave it OK, and she was mostly a lot more interested in just sort of the human behaviour learning. 175 00:19:03,850 --> 00:19:09,340 So that is the cut off. You can think of it in terms of MHR. We are done some of the work in those directions. 176 00:19:09,340 --> 00:19:18,160 Do we know, for example, you can do intention, recognition of the human model, 177 00:19:18,160 --> 00:19:24,040 basically vete in this particular case anymore to brain computer interface? 178 00:19:24,040 --> 00:19:28,450 This is nowhere near as fancy as the neural link that Elon Musk is trying. 179 00:19:28,450 --> 00:19:39,260 But even a simple thing which can work. Leslie know without actual explicit communication, there are what is actually able to see in this? 180 00:19:39,260 --> 00:19:46,300 The one in the lower window that that such in was the subject that actually wanted to reserve 181 00:19:46,300 --> 00:19:50,920 one of the blocks and that it realised that it was actually using only the other blocks. 182 00:19:50,920 --> 00:19:57,640 So things of this kind. So basically intention, recognition and anticipating what people want and wide and getting out of the way can be done. 183 00:19:57,640 --> 00:20:08,000 We have looked at that, you know, in the same direction, actually, robots can project their intentions in to the humans either by actually speaking. 184 00:20:08,000 --> 00:20:16,540 But in this particular case, we're using augmented reality holo lens, which sort of provides to that human not only what they're seeing, 185 00:20:16,540 --> 00:20:24,370 but also additional information that the robot is projecting about its goals by pointing out two particular blocks and so on. 186 00:20:24,370 --> 00:20:32,050 OK. So these are things that are possible. But I'm not actually going to be talking as much about that part. 187 00:20:32,050 --> 00:20:41,050 The one that I want to talk more about is essentially the fact that the moment humans are interacting with any AI system, 188 00:20:41,050 --> 00:20:46,690 they will wind up making their own model of daddy AI system. 189 00:20:46,690 --> 00:20:51,160 So if they see a robot, they will essentially have some model of the robot MRI. 190 00:20:51,160 --> 00:20:54,790 Which is the task model that they attribute to the robot. 191 00:20:54,790 --> 00:20:58,060 Some of this can be very magical, some of them, if they're experts in the loop. 192 00:20:58,060 --> 00:21:02,500 They might have some really good sense of what the work schools are, what it's actually capable of. 193 00:21:02,500 --> 00:21:12,460 So there is MRI edge, which is that humans model and the agent, the robot essentially now needs to have an approximation of that, too. 194 00:21:12,460 --> 00:21:19,900 This is the second mental model that I will spend most of my time in the dark talking about how to exploit and leverage that. 195 00:21:19,900 --> 00:21:29,350 So the MRI it here allows the agent to anticipate human expectations in order to either conform to those expectations. 196 00:21:29,350 --> 00:21:36,850 If it cannot confirm, then explain its own behaviour in terms by by sort of modifying a model. 197 00:21:36,850 --> 00:21:43,690 So tell the human you have to change your expectations on me in the following ways so that my behaviour makes sense to you. 198 00:21:43,690 --> 00:21:52,060 So to me, explanation really is this dialogue that the agent has with the human changing their model, which is their model. 199 00:21:52,060 --> 00:22:00,420 Updated. OK. I will mention that MHR and marriage are different from MH and Emaar. 200 00:22:00,420 --> 00:22:05,530 Imagine my father, human and father Robot Sackhoff execution models. 201 00:22:05,530 --> 00:22:10,360 These are models that they're hoping at executable in the world that actions without real preconditions. 202 00:22:10,360 --> 00:22:16,060 And, you know, hopefully, in fact, the application is satisfied. There should be a high probability that the action actually works. 203 00:22:16,060 --> 00:22:21,690 Unlike those image. Ah. And I'm on the edge of expectations. 204 00:22:21,690 --> 00:22:25,960 That really expectations on models. We mostly for this part of the stock. 205 00:22:25,960 --> 00:22:30,010 We'll talk about like the maximum likelihood expectations maybe. 206 00:22:30,010 --> 00:22:34,240 But in fact, they can even be like a distribution or what, the bardos. 207 00:22:34,240 --> 00:22:42,190 And they don't have to have any execution semantics at all. They are just being used by the agent to make sense of the other agents behaviour. 208 00:22:42,190 --> 00:22:53,320 So it's worth keeping that in mind. So given doors, it's basically the V, the longitudinal human A.I. Interaction Cycle works in its real assume. 209 00:22:53,320 --> 00:23:04,180 And I'm mostly focussing in this particular case in terms of the M.R. Edge, which is the human model of the robot's task model. 210 00:23:04,180 --> 00:23:08,530 And so you start with an initial estimate of haemorrhage. 211 00:23:08,530 --> 00:23:11,320 That is, the robot starts with the initial estimate of haemorrhage. 212 00:23:11,320 --> 00:23:16,930 And I should mention here that from my perspective, from any AI researchers perspective, 213 00:23:16,930 --> 00:23:23,740 if an agent is working with a human, the only controllable variable really is the eye agent. 214 00:23:23,740 --> 00:23:31,000 The most obviously controllable variable is the. And because you can't you can possibly tell the human, please do these. 215 00:23:31,000 --> 00:23:38,650 But they don't necessarily have to actually take your advice. So in essence, much of what I would be discussing is from the perspective of the robot. 216 00:23:38,650 --> 00:23:47,620 So not only does it needs to anticipate what the humans are thinking of for its capabilities, it also needs to. 217 00:23:47,620 --> 00:23:51,620 And I understand what the humans think of its capabilities. 218 00:23:51,620 --> 00:23:57,740 And I'm so used that that is our marriage and we assume that we start with our initial estimate of our marriage. 219 00:23:57,740 --> 00:24:02,120 And I talk to you in some cases, they're stocked with a shared model initially. 220 00:24:02,120 --> 00:24:05,720 Other cases you can actually learn from existing behaviour traces. 221 00:24:05,720 --> 00:24:10,160 I'll get to that part little later in the talk and then you sort of have this loop. 222 00:24:10,160 --> 00:24:12,350 There are two pieces in the loop. 223 00:24:12,350 --> 00:24:21,520 One is the model following where if the agent thinks that there is a sort of an understanding between what the human thinks its model is, 224 00:24:21,520 --> 00:24:28,250 that then it can just conform to it and marriage and conforming to it is really explicable. 225 00:24:28,250 --> 00:24:31,310 And I talk about it and I'll have it in a minute. 226 00:24:31,310 --> 00:24:39,860 So an agent basically making its behaviour such that it is in accordance with what a human agent expects it to do. 227 00:24:39,860 --> 00:24:48,200 That is explicable. There is another notion that other notions related to model following, which such as a predictability, 228 00:24:48,200 --> 00:24:54,380 which is sort of locally this behaviour should be consistent with what the human expects. 229 00:24:54,380 --> 00:25:05,280 Then that is the model. Communication is sometimes actually conforming to the human model, winds up being extremely costly. 230 00:25:05,280 --> 00:25:10,310 And so in those cases, essentially, that robot basically decides to do what is optimal for it. 231 00:25:10,310 --> 00:25:12,140 I'll give you an example of this in a minute. 232 00:25:12,140 --> 00:25:18,170 And then it realises upfront that since this behaviour is different from what the human would be expecting, 233 00:25:18,170 --> 00:25:24,010 given their model of its behaviour, it needs to change their model. 234 00:25:24,010 --> 00:25:31,550 And so that's what explicable the explanation part starts explanation really is communicating the changes to our marriage, 235 00:25:31,550 --> 00:25:37,910 either as it heads up it upfront or as a post facto explanation after the behaviour is done. 236 00:25:37,910 --> 00:25:47,660 There are other ideas such as legibility, which is sort of the explanation, except typically with implicit actions and design, 237 00:25:47,660 --> 00:25:54,190 which is sort of changing the environment in such a way that the environment itself gives you a heads up to the human. 238 00:25:54,190 --> 00:26:00,380 As to what you are. A marriage is also touch on a little bit on the design factor. 239 00:26:00,380 --> 00:26:08,690 So that's basically the longitudinal human interaction cycle. And I will sort of show this flash to slide a couple of times in the next few sites. 240 00:26:08,690 --> 00:26:16,100 So let's talk about model following part. So as I mentioned, I we actually use two kinds of use cases. 241 00:26:16,100 --> 00:26:21,200 One is on the left hand side where basically you have embodied robots. 242 00:26:21,200 --> 00:26:31,530 And in this particular case, where the robot, the Fitch robot, is going around in a mock search and rescue scenario with the human commander, 243 00:26:31,530 --> 00:26:39,590 are safely ensconced in some some other place and having like a camera, a dialogue with their robot on the other side. 244 00:26:39,590 --> 00:26:44,120 On the right side, we are looking at a planning decision, support system. 245 00:26:44,120 --> 00:26:49,490 If you have a mission plan, the Mars mission planner or somebody, they are trying to make the plans themselves. 246 00:26:49,490 --> 00:26:55,580 And the attention support system provides suggestions about how to complete their plans. 247 00:26:55,580 --> 00:27:03,560 So these two are the two general kinds of scenarios that we look at, the challenges that they look very different syntactically. 248 00:27:03,560 --> 00:27:11,870 What the challenge is about, how an agent interacts with the human wind up being quite similar in both cases as it is. 249 00:27:11,870 --> 00:27:15,320 I'll show you as we go forward. OK. Does other use cases. 250 00:27:15,320 --> 00:27:22,640 So and then, you know, given that I love social, mostly the robot use cases in the demos because those are more fun to watch. 251 00:27:22,640 --> 00:27:26,570 So in the case of model differences with the human in the loop, 252 00:27:26,570 --> 00:27:32,480 that means that emerick's that the human has is different from MRV, that the robots actual model. 253 00:27:32,480 --> 00:27:39,410 So that what that means is the plans that are optimal to the robot according to its model, 254 00:27:39,410 --> 00:27:48,140 Imod might actually not be optimal according to the human's expectations of a robot model. 255 00:27:48,140 --> 00:27:52,580 So I am on edge. So what the plan that is optimal in that model may not be optimal in commodities. 256 00:27:52,580 --> 00:27:53,840 That's all I'm saying. 257 00:27:53,840 --> 00:28:01,790 And when that happens, there's a surprise on the human side because they're basically seeing behaviour by the robot that they didn't expect. 258 00:28:01,790 --> 00:28:05,870 So this leads to inexplicability are inexplicable behaviour. 259 00:28:05,870 --> 00:28:14,810 At that point, the robot has two options. It can from in the model following is to change the model that the humans have. 260 00:28:14,810 --> 00:28:17,000 That is a marriage, which is the explanation part. 261 00:28:17,000 --> 00:28:23,510 And I'll talk to both of you, talk about both of them in the context of this search and rescue scenario. 262 00:28:23,510 --> 00:28:29,120 So imagine you have the you know, in this scenario, you have the human and the agent, 263 00:28:29,120 --> 00:28:34,190 the robot essentially start with the correct map of the environment in this particular case. 264 00:28:34,190 --> 00:28:40,370 This is the fifth floor right outside my door. And in computer science department right here is you. 265 00:28:40,370 --> 00:28:46,740 And presumably that is sort of a. So the model of the map of the environment is known upfront. 266 00:28:46,740 --> 00:28:50,250 But. It turns out that there is possibly some sort of a disaster. 267 00:28:50,250 --> 00:28:57,180 And so there possibly are rubble, there may be parts that are closed, maybe actually be passed out are opening up, 268 00:28:57,180 --> 00:29:04,350 which means that the Internet age and the robot, which is on the floor, might actually wind up realising that the map has changed. 269 00:29:04,350 --> 00:29:08,190 But the human still thinks that the map is the same as before. 270 00:29:08,190 --> 00:29:12,210 And so essentially, there's already a difference between marriage and a model. 271 00:29:12,210 --> 00:29:17,820 And then that needs to be reconciled through, you know, through this interaction process, 272 00:29:17,820 --> 00:29:24,150 either by the robot being conforming to what the human expects or by changing the model. 273 00:29:24,150 --> 00:29:30,720 So in the context of conforming to the human model in this particular case, in this picture, 274 00:29:30,720 --> 00:29:39,450 what's happening is that particular path that would have been the shortest path is closed because there is an obstacle act. 275 00:29:39,450 --> 00:29:46,440 Think of it as some kind of some huge double date. So if you don't want the human to be surprised at your behaviour because they're 276 00:29:46,440 --> 00:29:50,670 actually expecting perhaps that you would be coming out of this particular kind of dud, 277 00:29:50,670 --> 00:29:54,570 then you have to conform to the humans model. 278 00:29:54,570 --> 00:29:58,650 So in this particular case, it wound up removing the rubble. 279 00:29:58,650 --> 00:30:03,660 So it's costlier than just taking the next best path. That's what is explicable behaviour. 280 00:30:03,660 --> 00:30:06,900 So to be able to provide explicable security, given a goal, 281 00:30:06,900 --> 00:30:14,010 the objective normally would have been to just find a plan that is optimal with respect to its own model. 282 00:30:14,010 --> 00:30:17,910 But now it needs to find a plan that's not only opportunity inspectorates model, 283 00:30:17,910 --> 00:30:24,510 but it's close to what is expected by the human based on dead model, which is there at marriage. 284 00:30:24,510 --> 00:30:29,010 And so it becomes sort of a multi-modal planning problem. 285 00:30:29,010 --> 00:30:30,270 And then this can be done. 286 00:30:30,270 --> 00:30:40,620 I'll talk about this become sort of a more complex optimisation problem, you know, to actually come up with this kind of an explicable plan. 287 00:30:40,620 --> 00:30:44,430 And what we're trying to do is sort of reduce the distance as much as possible. 288 00:30:44,430 --> 00:30:52,110 Sometimes making the distance zero actually might be impossible because you must might be expecting magical properties that the robot doesn't have, 289 00:30:52,110 --> 00:30:57,150 such as completely removing the rubble from a particular candidate. 290 00:30:57,150 --> 00:31:01,500 OK. So now conforming to the explanations is costly. 291 00:31:01,500 --> 00:31:05,880 As I said something, you actually have to remove rubble or something in this particular example. 292 00:31:05,880 --> 00:31:13,170 So then you might want to communicate changes to am on edge to the human. 293 00:31:13,170 --> 00:31:18,230 And that would be the explanation. But in this case. 294 00:31:18,230 --> 00:31:25,640 If you could hear it, it was basically yelling at the top of the wires that a particular part is actually blocked. 295 00:31:25,640 --> 00:31:29,330 And because it's this other plot that is being taken. 296 00:31:29,330 --> 00:31:31,850 So in essence, it's just changed the model, 297 00:31:31,850 --> 00:31:39,050 just partially Delta changed to the model that the human and if the human actually pays attention and takes that into account, 298 00:31:39,050 --> 00:31:44,990 then all of a sudden the robot's behaviour is as explicable with respect to the changed model. 299 00:31:44,990 --> 00:31:53,720 So the explanations really has been talked about a lot in explainable A.I. systems and explanations have been talked about a lot. 300 00:31:53,720 --> 00:32:04,100 But when human A.I. agents are interacting, really it's best to think of explanations essentially as this constellation constellation of models. 301 00:32:04,100 --> 00:32:08,570 This is sort of close to psychological value here. You do as I talk about it later. 302 00:32:08,570 --> 00:32:12,620 But in essence, what we are seeing is that an explanation. 303 00:32:12,620 --> 00:32:25,970 Epsilon, given a plan by that robot, has essentially decided to act on is is such that if you add this Epsilon to current Modrich of the human, 304 00:32:25,970 --> 00:32:32,210 then this plan by will be actually optimal in this changed model. 305 00:32:32,210 --> 00:32:40,580 And it's also already optimised with respect to the robots model. That's why the robot again, I'm using the word optimal. 306 00:32:40,580 --> 00:32:47,300 It's easier to think in terms of essentially optimality from a theoretical analysis point of view. 307 00:32:47,300 --> 00:32:53,240 But in practise, essentially, you're more or less trying to say, I'm doing this and you're expecting a different plan. 308 00:32:53,240 --> 00:32:58,250 And this is the changes that I need to make to your model such that my plan would be in the Top Gear. 309 00:32:58,250 --> 00:33:04,090 Plans that you would see me doing about optimality allows for a much better theoretical analysis. 310 00:33:04,090 --> 00:33:08,660 OK, so that's the explanation. Size model reconciliation. It turns out that, in fact, 311 00:33:08,660 --> 00:33:13,160 computing these explanation's winds up being a metasearch in the space of the models 312 00:33:13,160 --> 00:33:20,190 themselves actually skipped a little bit about the representations of the models values. 313 00:33:20,190 --> 00:33:21,500 You know, in much of this, 314 00:33:21,500 --> 00:33:30,330 what we wound up using planning domain description languages which have actions and with the preconditions and effort in symbolic terms. 315 00:33:30,330 --> 00:33:37,580 And, um, so you if you have a model with a bunch of actions and the implications and effects of the model that the human has with a, you know, 316 00:33:37,580 --> 00:33:41,090 either similar odd different actions with their preconditions, 317 00:33:41,090 --> 00:33:50,030 what the robot is trying to do is essentially modify the human model, Emerich, that they are attributing to the robot. 318 00:33:50,030 --> 00:33:53,090 This will involve, for example, changing the preconditions, 319 00:33:53,090 --> 00:34:00,350 changing the effects of changing the costs of the actions that the humans have attributed to the robot. 320 00:34:00,350 --> 00:34:05,360 And this search essentially has all sorts of interesting properties. Using this model space search. 321 00:34:05,360 --> 00:34:07,970 You can come up with different kinds of explanations, 322 00:34:07,970 --> 00:34:15,380 both minimal explanations that would just make the human understand why this behaviour is fine in this case, 323 00:34:15,380 --> 00:34:22,910 but might question this explanation little later because you do something that's inconsistent with this explanation. 324 00:34:22,910 --> 00:34:28,670 You can also come up with monotonic explanations which will essentially not cross that confusion. 325 00:34:28,670 --> 00:34:31,670 Monotonic explanations will be costlier to compute. 326 00:34:31,670 --> 00:34:37,520 And one of the cases you are searching from the human model in the model space works, the robots model. 327 00:34:37,520 --> 00:34:46,980 In either case, you're searching from the other side. The technical details of this search are actually in this HK 2000 17 paper. 328 00:34:46,980 --> 00:34:53,780 That would be you might want to look at it, even though I sort of mentioned explicable. 329 00:34:53,780 --> 00:35:01,160 And explanation as if one is the model following and one is the model communication. 330 00:35:01,160 --> 00:35:09,950 You can actually combine both of them from a planning perspective, from a reasoning agent's perspective into a single planning framework. 331 00:35:09,950 --> 00:35:17,690 You know, we have a paper. Actually, I did this last year just before everything closed in play 20. 332 00:35:17,690 --> 00:35:27,500 There's a paper on expectation we're planning which essentially considers both on peak actions, the kinds of actions that change the environment. 333 00:35:27,500 --> 00:35:37,130 Mostly those are climate effects are changing the environment. Epistemic actions, which are mostly their effect, is to change the mental models. 334 00:35:37,130 --> 00:35:43,460 That means that communicative actions and sometimes you have actions and take actions can have epistemic effects. 335 00:35:43,460 --> 00:35:51,740 That's how you wind up doing implicit communication just by doing actions that have physical effects and then on rent. 336 00:35:51,740 --> 00:35:58,970 But they also are providing mental effects, a mental model changes to the human side. 337 00:35:58,970 --> 00:36:02,630 So the robot has both the standard Ontake actions and explanatory actions. 338 00:36:02,630 --> 00:36:06,980 It can model the effects of its actions on both states as well as the human's mental state. 339 00:36:06,980 --> 00:36:14,510 It sort of becomes for those of you and especially people like Mike. I mean, this is essentially epistemic planning, epistemic reasoning. 340 00:36:14,510 --> 00:36:17,620 So here, planning becomes a multi model and. 341 00:36:17,620 --> 00:36:25,540 Is it uses its both its model and the human's expectation model to generate a course of action that contains both explicitly unwanted actions. 342 00:36:25,540 --> 00:36:33,510 And so the explanatory actions are such that those are making changes to the marriage, that the human has an offer. 343 00:36:33,510 --> 00:36:42,220 Those changes are made. Then the rest of the Ontake actions that the plan that the robot is actually exhibiting, 344 00:36:42,220 --> 00:36:45,820 that behaviour is exhibiting would be optimal in this changed model. 345 00:36:45,820 --> 00:36:48,760 So you're doing both in the same environment. 346 00:36:48,760 --> 00:36:58,870 And the difference between applicability and explanation just becomes whether that's planetary message is a the E side pot is empty, 347 00:36:58,870 --> 00:37:01,360 in which case it's just explicable planning. 348 00:37:01,360 --> 00:37:07,270 And if it's not empty, then you are asking that you want to change that mental model first before understanding your behaviour. 349 00:37:07,270 --> 00:37:12,740 So that becomes explanation. But it turns out, actually epistemic planning has been studied before. 350 00:37:12,740 --> 00:37:23,680 Do one of the nice things is for just the two level nesting that we're considering that are efficient compilations to single agent planning. 351 00:37:23,680 --> 00:37:28,810 And so that's the trick that we use to make this doable, tractable. 352 00:37:28,810 --> 00:37:33,940 So the details of how the compilation works here are industry 2020. 353 00:37:33,940 --> 00:37:41,410 Although I looked at the robot changes, basically that the embodied robot scenarios, 354 00:37:41,410 --> 00:37:46,750 as I said, the same kinds of explanations can be used in the decision systems. 355 00:37:46,750 --> 00:37:56,020 So this is a radar system that we use to kind of provide vision, planning support to humans in the loop. 356 00:37:56,020 --> 00:38:02,920 We had some work at NASA as well as some building work. And now when our Office of Naval Research, in both cases, 357 00:38:02,920 --> 00:38:09,430 people are interested in making plans and people want to be in charge and the systems should be helping them, the A.I. system should be helping them. 358 00:38:09,430 --> 00:38:13,790 And the explanations, when you provide a solution, why did you provide this solution? 359 00:38:13,790 --> 00:38:16,780 If that human were to ask, you should be able to provide an explanation. 360 00:38:16,780 --> 00:38:22,250 And that is something that systems are able to do using these kinds of theories. 361 00:38:22,250 --> 00:38:27,370 OK, so coming back to the longitudinal human interaction cycle. 362 00:38:27,370 --> 00:38:35,980 Now, what I want to do is that having talked about Explicable Ety an explanation, I want to talk about the learning bot. 363 00:38:35,980 --> 00:38:42,910 So basically, where did that what did you want and what which model come. Sometimes it is as I said, it maybe did as a shared motivationally. 364 00:38:42,910 --> 00:38:50,950 Sometimes it has to be learnt. So in some cases, such as that USAF scenario human and the agent will both start with the same shared model. 365 00:38:50,950 --> 00:38:54,130 So all that is needed is tracking the model drift. 366 00:38:54,130 --> 00:39:01,320 Even if the robot doesn't know the model and Modrich with certainty, it can jism with multiple possible models. 367 00:39:01,320 --> 00:39:04,840 If you are not sure which is the specific model the robot a human has. 368 00:39:04,840 --> 00:39:09,120 If you don't know Emerich specifically, do you think that it's one of these scale models? 369 00:39:09,120 --> 00:39:13,630 You know, in the most general case, you have some Bayesian distribution of what these models, 370 00:39:13,630 --> 00:39:16,720 especially the case where you have kear different models. 371 00:39:16,720 --> 00:39:23,740 We can actually reason with these different models and provide conferment explanations of conditional explanations. 372 00:39:23,740 --> 00:39:30,130 With respect to these models, six models that there's a paper back in 2008 didn't come back. 373 00:39:30,130 --> 00:39:36,100 In other cases, in the end, you actually have to let alone these models from scratch. 374 00:39:36,100 --> 00:39:42,850 Some behaviour traces. If it's not being provided by upfront, by knowledge engineering aspects. 375 00:39:42,850 --> 00:39:47,830 So in that sense, the agent would need to learn the human and mental models from the traces. 376 00:39:47,830 --> 00:39:54,430 One thing that we do want to keep in mind is unlike MHR, which is anticipating humans, 377 00:39:54,430 --> 00:40:00,450 which can be learnt from human behaviour, traces that have been cached or at large amounts of time. 378 00:40:00,450 --> 00:40:04,660 You've just seen too many humans acting on the problem in over a long period of time. 379 00:40:04,660 --> 00:40:13,420 And you just use that odd edge which is requiring basic that is humans take on what the robot is doing. 380 00:40:13,420 --> 00:40:23,560 And so that doesn't just come from the human behaviour. Oftentimes we have different expectations of the agent than what we are capable of doing. 381 00:40:23,560 --> 00:40:29,120 So, for example, if you're walking to a room and you see a robot that starts dancing, 382 00:40:29,120 --> 00:40:37,210 you would be surprised because you don't expect robots to be able to dance, except maybe in a couple of, you know, videos that you may have seen. 383 00:40:37,210 --> 00:40:46,780 So whereas humans, you do expect them to be able to dance. So that MRI really requires places that human likes or the traces of the robot's behaviour. 384 00:40:46,780 --> 00:40:53,340 So that's a slightly harder learning problem. And sometimes you may actually have to deal with vocabulary differences, too. 385 00:40:53,340 --> 00:40:54,460 And I'll talk about that. 386 00:40:54,460 --> 00:41:08,810 I mean it with respect to one of the points that I want to mention here is that, as I mentioned, MHR and Edmar age are both expectations on models. 387 00:41:08,810 --> 00:41:17,230 And so they really don't have to be in any specific format, such as being little Fotomat that a modern MHR. 388 00:41:17,230 --> 00:41:21,520 Rewards. Because they actually have to plan with them. But that much out and am I right? 389 00:41:21,520 --> 00:41:26,280 I only SRF expectations and they can be represented in many more flexible ways. 390 00:41:26,280 --> 00:41:33,030 One of the things we did essentially is to learn human preferences of the robot behaviour as 391 00:41:33,030 --> 00:41:39,950 labelling functions during a training phase and use these labelling functions internally. 392 00:41:39,950 --> 00:41:43,350 But the robot then learns this labelling function. 393 00:41:43,350 --> 00:41:53,070 There's a paper in 2017 which sort of treats learns this as a sort of CRF and uses this labelling function as part of its behaviour, 394 00:41:53,070 --> 00:41:54,510 its planning process. 395 00:41:54,510 --> 00:42:05,370 So it computes this distance between the plan it is making and the plan the human expects in terms of this labelling procedure that it learnt. 396 00:42:05,370 --> 00:42:11,940 And so that is a way you can be explicable. It turns out that the same idea for learning. 397 00:42:11,940 --> 00:42:18,170 So you're essentially in this particular case, you're learning a marriage not by learning it in explicit actions, 398 00:42:18,170 --> 00:42:22,290 clear conditions affect state, but in terms of just a labelling procedure. 399 00:42:22,290 --> 00:42:27,780 You could do the same thing even when you are not doing explicable. But you're also providing explanations. 400 00:42:27,780 --> 00:42:32,310 In that case, all you need to do is improve the training scenario, 401 00:42:32,310 --> 00:42:37,890 expand the training scenario such that you are showing a behaviour as well as some explanatory 402 00:42:37,890 --> 00:42:44,580 messages next to them and asking you humans to say whether or not the behaviour makes sense. 403 00:42:44,580 --> 00:42:49,260 Which parts of the behaviour is making sense. Which doesn't. And this is the training phase. 404 00:42:49,260 --> 00:42:50,700 And having used this, 405 00:42:50,700 --> 00:43:01,080 then you can then have a labelling procedure which can then be used to compute the explanations on demand without having a an explicit EMERICH, 406 00:43:01,080 --> 00:43:07,890 which is in terms of actions, pretensions and effects. So this is why it's important to understand that Modrich and my child are just expectations. 407 00:43:07,890 --> 00:43:13,230 They don't have to be executable models. Having said all of this about explanation, 408 00:43:13,230 --> 00:43:21,490 some of you probably have heard a lot about explain ability and explain will I in the context of machine learning systems. 409 00:43:21,490 --> 00:43:26,770 I want to kind of make a connexion between all I discussed and some of the body of the work. 410 00:43:26,770 --> 00:43:38,310 They're both similarities and differences. And it's what understanding XY is hot, but mostly as a debugging tool for inscrutable representations. 411 00:43:38,310 --> 00:43:49,860 So, for example, oftentimes you lined up the system basically says this particular particular dog is an Alaskan husky. 412 00:43:49,860 --> 00:43:52,890 And you ask, why do you think it is an Alaskan husky? 413 00:43:52,890 --> 00:43:59,730 And the system points out that the snow part of the pig cells that the saliency region is, that's not big cells. 414 00:43:59,730 --> 00:44:03,330 And then you might say, aha, you don't really understand Alaskan husky at all. 415 00:44:03,330 --> 00:44:09,300 You're just understanding snow and the correlation of snow being present next to the dog. 416 00:44:09,300 --> 00:44:15,360 But notice, first of all, that this is a debugging explanation and it's a pointing explanation. 417 00:44:15,360 --> 00:44:19,310 And really, you can't actually point to many things. 418 00:44:19,310 --> 00:44:27,770 You know, oftentimes the kinds of behaviour I'm talking about, the sequential behaviour as to why did you do this decision? 419 00:44:27,770 --> 00:44:32,790 You you want to point out, you know, do a pointing explanation there. 420 00:44:32,790 --> 00:44:37,930 You need to point to a space time, too, which is extremely hard and very unwieldy, 421 00:44:37,930 --> 00:44:43,860 which is why civilisation progress by sort of developing these symbolic vocabularies through which we provide explanations. 422 00:44:43,860 --> 00:44:48,480 And that's the kind of model based explanations that I've been talking about till now. 423 00:44:48,480 --> 00:44:53,880 In fact, if you have looked at this adversarial example scenario that, you know, 424 00:44:53,880 --> 00:45:01,500 many people have probably seen about that school bus on the left hand side of a little bit of noise becomes an ostrich on the right side, 425 00:45:01,500 --> 00:45:08,490 you and you don't see it as an ostrich. You know, our system, a classification system might very well say with high confidence that it is an ostrich. 426 00:45:08,490 --> 00:45:14,250 And if you ask it, tell me which part of this makes it. Moscovitch, point to me what parts makes an ostrich. 427 00:45:14,250 --> 00:45:20,510 That would be a useless explanation because it'll just show a couple of several pixels, 428 00:45:20,510 --> 00:45:25,560 you know, all over the picture, which don't have any rhyme or reason from your point of view. 429 00:45:25,560 --> 00:45:30,090 But that's what the system is concerned that made that class for this on ostrich. 430 00:45:30,090 --> 00:45:34,470 So the point being that pointing explanations of the primitive explanations that 431 00:45:34,470 --> 00:45:39,390 are mostly useful only for debugging explanations are critical for collaboration, 432 00:45:39,390 --> 00:45:47,140 but they're not really a solid luckly by the agent. It's not just talking to itself, which is trying to make sense to the human agent in the loop. 433 00:45:47,140 --> 00:45:50,430 And that would require a lot more than pointing pixels. 434 00:45:50,430 --> 00:45:57,420 And the model, the Constellation view that I've been talking about here views closer to psychological theories of explanation, 435 00:45:57,420 --> 00:46:01,410 such as the ones that not Labruzzo uncle talked about. 436 00:46:01,410 --> 00:46:15,270 That's a useful thing to keep in mind. So. One other thing I want to mention when I'm talking about here is that handling differing vocabulary's. 437 00:46:15,270 --> 00:46:18,030 Is another important issue that could wind up happening. 438 00:46:18,030 --> 00:46:25,130 We know we did not talk about whether the models are on completely different vocabularies at modern Modrich. 439 00:46:25,130 --> 00:46:31,120 Then you have an inscrutable system which essentially, basically is learning its own representations. 440 00:46:31,120 --> 00:46:37,310 And so there is no direct connexion to any concepts that the human understands. 441 00:46:37,310 --> 00:46:40,950 This is where you were hoping that you'd be able to just point to the pixels of 442 00:46:40,950 --> 00:46:44,930 something that is common between us and point to that particular substrate. 443 00:46:44,930 --> 00:46:48,170 But really, if you have things like windshear vision problems, 444 00:46:48,170 --> 00:46:53,600 you will need to provide the explanation in terms of concepts that humans can understand. 445 00:46:53,600 --> 00:47:01,890 And so ongoing work that we have in a lab, as well as some other work that some other people like Kim have done, involves essentially. 446 00:47:01,890 --> 00:47:08,520 Translating the explanation that you have into the vocabulary that the humans understand and to do that translation, 447 00:47:08,520 --> 00:47:14,460 you need to first learn the translation so you understand mappings from simple to big cells in this particular case. 448 00:47:14,460 --> 00:47:20,250 And then can work, can hold the explanation in to these symbols. 449 00:47:20,250 --> 00:47:23,430 And when you do that, it will be possibly an approximate explanation. 450 00:47:23,430 --> 00:47:31,830 But that's way better than just sort of throwing a big place off your reasoning to the humans and say this is the reason I did it. 451 00:47:31,830 --> 00:47:36,260 Actually, I should mention that that's trivial. Violist farm off explanations. 452 00:47:36,260 --> 00:47:40,890 Are those where you say you will understand what I did. If you take my brain here, take it. 453 00:47:40,890 --> 00:47:46,980 And that is one of the cubicle farms. And we understand we assume that that that's not a good explanation at all. 454 00:47:46,980 --> 00:47:51,840 That's why we actually search for smaller pieces of changes to the marriage. 455 00:47:51,840 --> 00:48:00,440 That will be enough that human to understand your decisions, just to put this in a classification kind of scenario. 456 00:48:00,440 --> 00:48:06,210 We have some recent work which shows that sort of thinking about mental models, 457 00:48:06,210 --> 00:48:09,810 even if it's not sequential additions, making problems can be quite useful. 458 00:48:09,810 --> 00:48:16,440 So, for example, in many classification systems are typically evaluated only in terms of their accuracy. 459 00:48:16,440 --> 00:48:25,830 How often? Not correct. But I would argue that really it's equally important to see when they are wrong, how badly off are they? 460 00:48:25,830 --> 00:48:31,650 So a system which makes an egregious misclassification. Humans, we lose trust in it. 461 00:48:31,650 --> 00:48:35,440 By the way, the word plussed is the first time I mentioned it in the dog, but. 462 00:48:35,440 --> 00:48:40,670 The interaction between an agent and a human humans might kind of, you know, 463 00:48:40,670 --> 00:48:46,130 might engender trust in humans if in fact they can make sense of the decisions that the agent is making. 464 00:48:46,130 --> 00:48:49,180 So in this particular pictures that they're showing, 465 00:48:49,180 --> 00:48:55,040 we essentially about this system where it not only takes into account the classification accuracy, 466 00:48:55,040 --> 00:48:59,210 but that misclassification egregiousness and using that, 467 00:48:59,210 --> 00:49:06,320 you know, our system, for example, will be able to come up with better kinds of these classifications, 468 00:49:06,320 --> 00:49:10,070 even if it is asked to misclassify than the existing system. 469 00:49:10,070 --> 00:49:21,520 So, for example, some of the examples there are showing that, you know, if you have something like a Dabby on the very first top left gardener, 470 00:49:21,520 --> 00:49:27,200 you know, existing systems might just when they make a mistake with this, they might just say it's a remote control. 471 00:49:27,200 --> 00:49:34,650 And, you know, our system will say it's a tiger cat. So because in essence, it's also looking for the cost of that misclassification. 472 00:49:34,650 --> 00:49:42,560 And that's something that you can do with respect to mental models. So there's a bunch of things that are that can be extended from this. 473 00:49:42,560 --> 00:49:44,870 For example, communicating the model. 474 00:49:44,870 --> 00:49:53,150 Myard may not guarantee portability unless the human has the inferential capacity to compute optimal behaviour from the model. 475 00:49:53,150 --> 00:49:56,120 I sort of, you know, pushed out of the bag earlier. 476 00:49:56,120 --> 00:50:03,290 It turns out that you can use file based interactions with a bunch of what that we have done their way into humans. 477 00:50:03,290 --> 00:50:05,950 Just ask why not this other behaviour? 478 00:50:05,950 --> 00:50:12,620 How do you only provide reason as to why that other behaviour is not as good as the behaviour you have sheets shown. 479 00:50:12,620 --> 00:50:18,020 And this will be easier start off to understand than just dealt us to the model with 480 00:50:18,020 --> 00:50:22,930 respect to which you need to be able to see that the behaviour is actually optimal. 481 00:50:22,930 --> 00:50:26,840 And similarly, when you provide information to the humans, 482 00:50:26,840 --> 00:50:30,860 the real question is also a question is also whether the agent is actually 483 00:50:30,860 --> 00:50:35,060 humans are paying attention to the communication and that involves two parts. 484 00:50:35,060 --> 00:50:38,260 Is the communication actually perceivable, at least in theory, at least? 485 00:50:38,260 --> 00:50:46,520 So, for example, if I was getting this talk on radio talk radio, if I started showing you slides, holding them in my hands, 486 00:50:46,520 --> 00:50:53,780 the joke is on me because I'm not reasoning about the fact that you won't be able to see what I am sure you. 487 00:50:53,780 --> 00:50:59,090 So that is basically reasoning with the communication, possible perception modalities. 488 00:50:59,090 --> 00:51:03,920 The other is, even if you see, you may not be attention to it. 489 00:51:03,920 --> 00:51:07,430 That is attention management. That is something we haven't done much. 490 00:51:07,430 --> 00:51:16,220 But in fact, it's been done by others, such as the incorporates know in the context of the all Clippy, the Microsoft assistant, 491 00:51:16,220 --> 00:51:21,680 which is basically figuring out what is the best time to tell humans anything so that they're likely to pay attention to it. 492 00:51:21,680 --> 00:51:28,440 So in terms of the file based explanations, for example, we can you know, here's a bunch of papers that we have made. 493 00:51:28,440 --> 00:51:35,540 The humans basically ask, why not this behaviour? That is, they'll provide a plan, a partial plan of the company then and say, 494 00:51:35,540 --> 00:51:39,800 why aren't you doing this one against the one second is the one you are currently doing. 495 00:51:39,800 --> 00:51:46,330 And you can use that question itself to figure out what level of abstraction and the abstraction hierarchy 496 00:51:46,330 --> 00:51:52,220 of models that human side in and use that knowledge to provide that kind of explanations to them. 497 00:51:52,220 --> 00:52:01,030 That's like a file based explanation. We actually did this in the context of this sub radar system that we're talking about, 498 00:52:01,030 --> 00:52:08,120 where we provide a longitudinal dialogue between the agent and the humans and the human actually provides the file. 499 00:52:08,120 --> 00:52:12,200 Not only can we provide an explanation as to why that file won't work, 500 00:52:12,200 --> 00:52:19,070 but we also realise that that question means that the humans really would like the plan to be closer to the file. 501 00:52:19,070 --> 00:52:27,140 So then we do have a planning scenario where we will bind up coming up with a plan that's closer to the file, even if it is a little less optimal. 502 00:52:27,140 --> 00:52:32,990 And there's a demo CHIPLEY. I know that that talks about this. 503 00:52:32,990 --> 00:52:37,970 We also extended these models to also consider multiple humans in the loop. 504 00:52:37,970 --> 00:52:45,650 So essentially providing customised explanations for different humans at different times based on their mental models. 505 00:52:45,650 --> 00:52:51,380 And then I talked about this controllability of zero controlled observability planning problem. 506 00:52:51,380 --> 00:52:55,070 This is something that Unida has done there. Essentially, 507 00:52:55,070 --> 00:53:04,490 the system can reason about what can and cannot be seen by the human and use that both to make sure that they see that the help is being given. 508 00:53:04,490 --> 00:53:06,280 And also, if it if it's so, please, 509 00:53:06,280 --> 00:53:14,300 it's up Outfest gate to people in the loop for it doesn't want them to understand that this help, it can obfuscate. 510 00:53:14,300 --> 00:53:19,250 So that's also possible. That's actually a 19 paper. 511 00:53:19,250 --> 00:53:21,000 And then more recently, 512 00:53:21,000 --> 00:53:28,910 it has been looking at this proactive assistance part where essentially you are reasoning about the fact that help should not only be given, 513 00:53:28,910 --> 00:53:34,940 but should be seen to be given. So at least you should be able to reason about the possible upsetter. 514 00:53:34,940 --> 00:53:40,610 Aspects of the model observability aspects and make sure that the human understands 515 00:53:40,610 --> 00:53:45,470 that you have provided to help either by explicit communication in Oneida's work, 516 00:53:45,470 --> 00:53:52,670 it's actually by implicit communication. So you wind up showing that you are transporting second objects and that will sort of, 517 00:53:52,670 --> 00:53:58,490 you know, provide to communicate certain messages to the humans. 518 00:53:58,490 --> 00:54:04,670 So I just want to quickly go what this and maybe in the couple of minutes, I know that they're very close to hand. 519 00:54:04,670 --> 00:54:08,540 So the framework for human interaction really cannot just be me. 520 00:54:08,540 --> 00:54:12,320 I'm saying I did what because we are humans and we think it works for us. 521 00:54:12,320 --> 00:54:18,940 So our solution for this has been interdisciplinary collaboration with the human factors such as Nancy Cook. 522 00:54:18,940 --> 00:54:25,850 Sean Did is a colleague here, and she has actually been she was a past president of the Human Factor Society. 523 00:54:25,850 --> 00:54:31,160 And so we work with hard and hard group in setting up systematic human subjects, 524 00:54:31,160 --> 00:54:37,370 studies to actually cheque whether the kinds of explicable behaviour as well as explanations 525 00:54:37,370 --> 00:54:43,810 actually are coming up with makes sense to realise humans in the loop and doing human subjects. 526 00:54:43,810 --> 00:54:52,100 That is pretty painful. But I would also say that anybody who says that into human interaction but haven't heard of the word I ought to be, 527 00:54:52,100 --> 00:54:54,650 which is institutional review board certification, 528 00:54:54,650 --> 00:55:00,350 I'm probably making it all up because without that, you have no clue as to whether or not these things are actually working. 529 00:55:00,350 --> 00:55:02,960 So, in fact, there's papers in both H and I. 530 00:55:02,960 --> 00:55:16,770 And it's a journal that show how that kind of explanation and that we come up with are seen to be useful in collaborative behaviour with humans. 531 00:55:16,770 --> 00:55:22,430 A very lasting mental model capabilities allow agents to manipulate the humans in the loop. 532 00:55:22,430 --> 00:55:25,460 As I mentioned, they can allow for head fakes. 533 00:55:25,460 --> 00:55:33,620 This sort of leads to a bunch of ethical quandaries far any work which starts thinking in loving human agent agents, 534 00:55:33,620 --> 00:55:41,090 true mental model humans evolutionarily mental modelling allowed us to both cooperate as well as compete about each other. 535 00:55:41,090 --> 00:55:48,260 But really, you can't just stop your agents from having that ability because without the mental model, they won't cooperate with you. 536 00:55:48,260 --> 00:55:56,030 Even so, you will have to essentially deal with the Pandora's box like that has to be open you and systems with mental more linkable. 537 00:55:56,030 --> 00:56:01,640 It is being ethical boundaries, additional ethical quandaries beyond the usual. 538 00:56:01,640 --> 00:56:04,220 Oh my God. Autonomous robots are going to kill us all. 539 00:56:04,220 --> 00:56:10,620 Such as, for example, automated negotiating agents that misrepresent their intentions to gain material advantage. 540 00:56:10,620 --> 00:56:16,150 Your personal assistant that tells you white lies to get you to eat healthy. 541 00:56:16,150 --> 00:56:22,220 I end up, as I already mentioned, people would have been actually more guarded with social media if they realised that those platforms that 542 00:56:22,220 --> 00:56:28,220 actually actively profiling that we are already unmoored off our guard when we're dealing with other humans. 543 00:56:28,220 --> 00:56:32,360 But sometimes we don't realise that Facebook is actually an agent. 544 00:56:32,360 --> 00:56:36,920 It just doesn't look like a human. But it has some of the same capabilities in terms of profiling. 545 00:56:36,920 --> 00:56:44,640 So, you know, in particular that with the social intelligence scenarios, humans example, closer sentences are far more pronounced. 546 00:56:44,640 --> 00:56:49,740 And so we need to be very careful about how to deploy these technologies. 547 00:56:49,740 --> 00:56:55,160 As I said, head fakes really are a lot more powerful than deep fakes, and we don't need to worry about them. 548 00:56:55,160 --> 00:56:58,970 But then again, every tool is actually a weapon, too. If you just hold it right. 549 00:56:58,970 --> 00:57:01,420 And that's true also for these kinds of things. 550 00:57:01,420 --> 00:57:09,520 You know what what we actually have shown that model, the Constellation, can be used to tell lies, lies of commission as well as less commission. 551 00:57:09,520 --> 00:57:20,350 And then this work by Unagi that I mentioned earlier, that actually shows, for example, that that. 552 00:57:20,350 --> 00:57:25,990 In this particular case, in one case, the robot is making sure that the guy sees what it is doing. 553 00:57:25,990 --> 00:57:31,340 Watch what the robot is doing. Novik, is it actually dyspraxia to do something else? 554 00:57:31,340 --> 00:57:39,040 And so basically it's about manipulating the perception model. And so this is like signalling France and taking to enemies simultaneously. 555 00:57:39,040 --> 00:57:45,430 These are things that can be done. The purpose on this in 2009 interview a monster 10 20. 556 00:57:45,430 --> 00:57:53,830 Finally, you know what we did when I mentioned that lying robot study saying when can I box like we started under what conditions? 557 00:57:53,830 --> 00:57:58,900 People are willing to be told how light lies. And there's a paper on that. 558 00:57:58,900 --> 00:58:03,480 Yes. System conference. So I'll stop here. 559 00:58:03,480 --> 00:58:07,660 So somebody basically what I was trying to get to you is that effective human 560 00:58:07,660 --> 00:58:12,400 interaction requires human systems to be able to manage your own mental models. 561 00:58:12,400 --> 00:58:16,690 But the model that you want to model up their task as well as the humans model of the robot, 562 00:58:16,690 --> 00:58:20,570 at the least you could do crazy things with infinite regress. 563 00:58:20,570 --> 00:58:24,550 But in fact, you can do a lot of useful things with just the second level models. 564 00:58:24,550 --> 00:58:30,730 That's what I've tried to tell you. Managing mental models brings up inference as well as learning challenges. 565 00:58:30,730 --> 00:58:35,740 And these frameworks for human interaction have to be evaluated with actual human subjects. 566 00:58:35,740 --> 00:58:41,290 Studies and then mental modelling capabilities do open up new ethical problems. 567 00:58:41,290 --> 00:58:48,160 And we have to be careful about them. There's a paper in my magazine which sort of gives an overview of some of these issues. 568 00:58:48,160 --> 00:58:56,600 And then we also just wrote sort of a draught of a book that's under review, not explainable human interaction, planning perspective. 569 00:58:56,600 --> 00:59:07,440 With that, I'm going to stop and thank you for your attention. 570 00:59:07,440 --> 00:59:16,770 Thank you so much for that, rather, it was fantastic. So we are ready for questions, so please enter some questions in the Q&A. 571 00:59:16,770 --> 00:59:24,620 There is if you look on your teams, there is a button on the top right foot, which is cute. 572 00:59:24,620 --> 00:59:28,650 And there is a little question mark of speech bubble with a question mark on. 573 00:59:28,650 --> 00:59:33,850 If you click on that, that will get you to the questions and answers. Let me kick things off. 574 00:59:33,850 --> 00:59:38,910 So I absolutely see where you're coming from or where you want to go with this. 575 00:59:38,910 --> 00:59:46,800 But how far can you get with this towards, you know, to kind of human level modelling? 576 00:59:46,800 --> 00:59:50,670 How far will this take you? Otherwise, there's something that's missing. What do you think? 577 00:59:50,670 --> 00:59:54,780 This is all that we're going to need? Meaning what? Human level modelling. 578 00:59:54,780 --> 00:59:58,290 And this is the way we deal with each other. Yeah, yeah. 579 00:59:58,290 --> 01:00:08,850 I mean, this is partial steps towards it. And it's very useful in actual quotidian interactions between human and A.I. systems. 580 01:00:08,850 --> 01:00:18,010 I do realise that we do allow far sort of large amount of nesting when we design with each other. 581 01:00:18,010 --> 01:00:19,410 You know, oftentimes, in fact, 582 01:00:19,410 --> 01:00:27,510 one of the interesting things which I didn't talk about it is you can engender trust in somebody by doing what they expect you to do. 583 01:00:27,510 --> 01:00:30,690 And once they start trusting you, you can actually use it. 584 01:00:30,690 --> 01:00:38,250 You are at one day. So, in fact, we have a paper where this sort of reason about how much trust the labour written or robot should engender. 585 01:00:38,250 --> 01:00:45,330 So that it's not required to be explicable in the sense are not being monitored as closely. 586 01:00:45,330 --> 01:00:51,270 So that it turns out that given that people understand this already, you know, when you and I interact, 587 01:00:51,270 --> 01:00:57,140 if you always do things that I expect you to do, I start wondering, you know, stuff not being surprised. 588 01:00:57,140 --> 01:01:04,380 I mean, in the beginning, I'm not surprised because you're doing what I expecting you to do, but you're always doing what I'm expecting you to do. 589 01:01:04,380 --> 01:01:14,250 Then I wonder that you how other al-Qaeda goals and then so that will basically bring in additional nesting off the mental models, 590 01:01:14,250 --> 01:01:21,750 which we don't deal with right now. I would suddenly think that is something that will be important in bringing it closer. 591 01:01:21,750 --> 01:01:27,240 The other thing, of course, is really how we wound up human to human communication. 592 01:01:27,240 --> 01:01:35,610 We wound up developing this sort of a common shared vocabulary, you know, over a period of evolution as well as our number going up. 593 01:01:35,610 --> 01:01:39,780 And this is something that was very important. 594 01:01:39,780 --> 01:01:46,440 Far, far. Yeah, yeah. Agents in particular, those which are learning their own representations, 595 01:01:46,440 --> 01:01:51,030 in particular because the representation they learn may not have any connexions to what humans understand. 596 01:01:51,030 --> 01:01:56,220 And so this idea of translation, this is something that's going to be quite important. 597 01:01:56,220 --> 01:02:02,820 So I don't really know if this is still a big issue as to whether my thinking is in the same words that I'm talking to. 598 01:02:02,820 --> 01:02:07,710 You are I'm thinking in a different way and I'm talking to you in words that you will understand. 599 01:02:07,710 --> 01:02:10,950 And this is something that the agents have to deal with, like more. 600 01:02:10,950 --> 01:02:19,010 And this translation problem is something that I mentioned some of it earlier, but that's going to be another big issue in getting there. 601 01:02:19,010 --> 01:02:23,240 OK. So we have a question that's come in from Max, rankly. Max. 602 01:02:23,240 --> 01:02:24,720 Yes. Thank you for a brilliant lecture. 603 01:02:24,720 --> 01:02:31,020 How is this approach similar to the ways that intelligent tutoring systems model learning skills and understandings 604 01:02:31,020 --> 01:02:37,440 and then came to lessons and follow on skills reinforcement appropriately other existing similarities? 605 01:02:37,440 --> 01:02:42,870 And you think your approach will enable intelligent tutoring systems to be better tutors? 606 01:02:42,870 --> 01:02:49,800 Yeah, that's very much online. So in fact, I should have mentioned that when I said human interaction, 607 01:02:49,800 --> 01:02:57,930 I didn't quite explain who is helping whom and who is trying to get out of who's way in the way we do the work. 608 01:02:57,930 --> 01:03:01,470 Actually, the robot could be the teacher, the robot could be the student. 609 01:03:01,470 --> 01:03:08,160 That or what could be peer to peer and ideas basically is the robot as the tutor to the human. 610 01:03:08,160 --> 01:03:15,120 And in many of these issues, actually, in fact, as I mentioned in the very beginning on the human eye, very AI application slide, 611 01:03:15,120 --> 01:03:22,440 people who have always given let's speak to humans in the loop are intelligent tutoring systems, people who like out on land. 612 01:03:22,440 --> 01:03:30,930 My colleague in that field here and other ideas, people, though, always that they didn't have the luxury of saying let's ignore humans, basically. 613 01:03:30,930 --> 01:03:33,180 The whole point is trying to teach the humans. 614 01:03:33,180 --> 01:03:40,860 And so, yes, there's a very significant connexion between the ideas of literature and the work that we are doing. 615 01:03:40,860 --> 01:03:50,190 We actually have papers on applying some of this model reconciliation framework in teaching scenarios that, 616 01:03:50,190 --> 01:03:55,080 you know, this is an AI caps workshop paper that's available from our homepage. 617 01:03:55,080 --> 01:03:58,320 In fact, one of my students, our current students, Suchin, 618 01:03:58,320 --> 01:04:03,510 was actually in one of the earlier videos where he tries to get away from the robot hand he worked with Good. 619 01:04:03,510 --> 01:04:12,960 And he's working with me. And so there is this obvious kind of. That idea systems and that human interaction are very much connected. 620 01:04:12,960 --> 01:04:21,270 One typical issue has been that idea systems tended to be sort of more samite automated oftentimes. 621 01:04:21,270 --> 01:04:30,570 And the level of explanations that they provide are somewhat different from the kinds of explanations we tended to provide in the what. 622 01:04:30,570 --> 01:04:34,800 But they're very much complementary and it would be less than that. 623 01:04:34,800 --> 01:04:38,580 We try to learn from the ideas, literature and also fought back. Yeah. 624 01:04:38,580 --> 01:04:49,240 And suddenly that's a clear cut case. In fact, I why has this new project coming up got into a perception guided task training? 625 01:04:49,240 --> 01:04:53,280 That's something where the robot will teach humans how to do a factory. 626 01:04:53,280 --> 01:04:58,890 What does any agent teach factory worker? How to do new physical activities by example. 627 01:04:58,890 --> 01:05:03,560 So it's a robot coach and that winds up having all these issues. 628 01:05:03,560 --> 01:05:11,300 You know, you need to have a marriage. You need to also have event shot and use them to provide these kinds of support. 629 01:05:11,300 --> 01:05:15,630 Okay. So. Well. So everybody out there please ask your questions. 630 01:05:15,630 --> 01:05:19,830 We've got another few minutes while we're waiting Sittler. Any other questions. So one other one for me. 631 01:05:19,830 --> 01:05:26,250 So so I was really pleased to see you allude to the work on the social brain hypothesis, 632 01:05:26,250 --> 01:05:32,440 which is this idea that we have big brains, which enables everything because we have large social groups. 633 01:05:32,440 --> 01:05:39,870 So we need to keep track of complex social relationships. And the famous work there, I think is work by Robin Dunbar at Oxford. 634 01:05:39,870 --> 01:05:43,980 Evolutionary psychologist, which is lovely, lovely work. 635 01:05:43,980 --> 01:05:49,590 It's the kind of work that I look at and think, well, I wish I did that kind of research. You know, that's. 636 01:05:49,590 --> 01:05:51,030 So what do I think about that work? 637 01:05:51,030 --> 01:05:59,430 I mean, the the example that I use to illustrate it is this six word dialogue that's due to Steven Pinker where where Bob says, 638 01:05:59,430 --> 01:06:04,590 I'm leaving you and says, who is she? Right. And I love it because it's so crisp. 639 01:06:04,590 --> 01:06:09,540 Six words. And again, it's incredibly rich. And we all understand what's going on there. 640 01:06:09,540 --> 01:06:15,450 And we can all fill in all the details about what's going on in the relationship between, you know, 641 01:06:15,450 --> 01:06:22,110 we can draw rich, rich, rich pictures of these two individuals based on just those six words. 642 01:06:22,110 --> 01:06:28,200 And that a very important part of that understanding is our kind of common sense theory of mind. 643 01:06:28,200 --> 01:06:38,070 We all have. So what I'm puzzling over and this is entirely independently of your view of your talk is is it a target? 644 01:06:38,070 --> 01:06:43,200 Should it be a target of a A.I. to build machines that can understand that dialogue? 645 01:06:43,200 --> 01:06:51,960 I mean, you imagined as a robot in the room with with ball and should it understand should what should it do modelling and understand that dialogue? 646 01:06:51,960 --> 01:06:56,040 I think it's certainly my say. I said basic. 647 01:06:56,040 --> 01:07:03,050 I mean, so what would the worry, presumably, that you have is that it might lead to other unethical uses. 648 01:07:03,050 --> 01:07:08,520 I mean, I think that ship has sailed my sentences and everything. 649 01:07:08,520 --> 01:07:14,370 I mean, that's what I think. I mean, every technology is dual use technology and intelligence is the ultimate technology. 650 01:07:14,370 --> 01:07:19,080 So everything we do in the context of intelligence is going to be dual use. 651 01:07:19,080 --> 01:07:24,330 You cannot stop developing it. You have to dial up with complete care. 652 01:07:24,330 --> 01:07:32,040 They can to make sure that, you know, some some sorts of ethical framework is being sort of followed, even flying them. 653 01:07:32,040 --> 01:07:38,400 So my science and my sense of as an AI problem, as a problem of not understanding intelligent interaction. 654 01:07:38,400 --> 01:07:43,050 This is very much relevant to actually fill up these blanks. 655 01:07:43,050 --> 01:07:46,860 I mean, you know, using the significant amount of common sense background knowledge, 656 01:07:46,860 --> 01:07:55,110 what actually has happened as to how those sorts of things lined up, becoming a problem, of course. 657 01:07:55,110 --> 01:07:59,280 Everything I talked about almost can be misused. 658 01:07:59,280 --> 01:08:04,300 And so that's that's just the issue. 659 01:08:04,300 --> 01:08:09,210 Yeah. So I would say it suddenly is what worth doing and in fact, part of the idea. 660 01:08:09,210 --> 01:08:14,070 They will be dealing with inferences in the context of rich background knowledge. 661 01:08:14,070 --> 01:08:17,800 I've sort of made it simpler by saying the model only with respect to the model. 662 01:08:17,800 --> 01:08:23,750 The inference is being done, but model plus everything else you know about the word will come into the picture. 663 01:08:23,750 --> 01:08:27,750 Doing those influences. OK. OK. So we go. 664 01:08:27,750 --> 01:08:35,390 I think one last question. So the question is human working relationships require time to establish after a few months with a new colleague. 665 01:08:35,390 --> 01:08:39,770 One starts to learn whether competence can be relied upon. I think there's a backstory there. 666 01:08:39,770 --> 01:08:41,870 And for example, jargon develops. 667 01:08:41,870 --> 01:08:50,570 To what degree are your eyes learning about individuals of different categories of human behaviour and building these relationships over time? 668 01:08:50,570 --> 01:09:01,550 So that's a fantastic question. So as I mentioned, this one slide about longitudinal picture of model following and model communication. 669 01:09:01,550 --> 01:09:09,810 This is something that, you know, I've sort of used to motivate a lot of our work, more stuff or beginning stuff. 670 01:09:09,810 --> 01:09:15,470 I work up until, let's say, last two years back was in the fast shot explanation. 671 01:09:15,470 --> 01:09:25,880 But longitudinal explanations lined up being changing your mental model of the human as well as JD noticing that data on commodities changing. 672 01:09:25,880 --> 01:09:35,150 A classic example is this idea that if the robot is doing something inexplicable in front of it, who won the first time? 673 01:09:35,150 --> 01:09:39,770 They'll be surprised. The third time they will still be supplies by 17th time. 674 01:09:39,770 --> 01:09:44,390 The have nots are placed on by the 18th time and the robot actually realises that it's making 675 01:09:44,390 --> 01:09:48,350 a mistake and starts doing the more explicable thing that humans do against uprights, 676 01:09:48,350 --> 01:09:54,890 because by this time they went around changing their model of how the human robot marriage. 677 01:09:54,890 --> 01:09:58,340 And this is something that happens in longitudinal scenarios. 678 01:09:58,340 --> 01:10:05,000 In that cluster work that I mentioned, we are actually trying to use that to see that if you engender enough trust, 679 01:10:05,000 --> 01:10:11,270 then you will basically not how to continually show explicable behaviour. 680 01:10:11,270 --> 01:10:16,520 So if you're shown enough explicable behaviour, some number of times, humans, basically how I trust them. 681 01:10:16,520 --> 01:10:23,640 So they have less of a monitoring going on on you, at which point you can live your life and do more things that are more optimal. 682 01:10:23,640 --> 01:10:28,160 Again, when you're doing this, whether in fact that's going to be unsafe or not, that's it. 683 01:10:28,160 --> 01:10:33,830 Actual separate issues or whether that robot is doing something deliberately unsafe with a human. 684 01:10:33,830 --> 01:10:39,820 Is that something that we have to still deal with? 685 01:10:39,820 --> 01:10:47,810 No, but the idea that our mental models will be changing during the interaction and so explanations 686 01:10:47,810 --> 01:10:52,820 will be changing as the mental models change is very much part of the work that we are doing. 687 01:10:52,820 --> 01:10:57,020 The more recent papers, in fact, is one that I didn't get into, 688 01:10:57,020 --> 01:11:07,820 which talks in terms of a Bayesian account of all these things where you think of what the human essentially has at the Bayesian prior or what a 689 01:11:07,820 --> 01:11:11,750 marriage that is a D of a possible different models that are what has and then 690 01:11:11,750 --> 01:11:19,340 Ventilla that humans are looking at what the robot is doing effectively, that the prior is being updated to be posted here. 691 01:11:19,340 --> 01:11:23,150 And so there are new probability masses, all of this. 692 01:11:23,150 --> 01:11:29,990 And in fact, there's always this possibility that one of the models that the humans have is, oh, my God, I have no idea what the robot is doing, 693 01:11:29,990 --> 01:11:37,370 which is what we call this Amistad, which means this is you admit to yourself that I have no clue what this person is doing. 694 01:11:37,370 --> 01:11:42,590 And I just hope that that would be the least likely explanations for the behaviour. 695 01:11:42,590 --> 01:11:48,650 And the more that becomes the most likely explanation, the more the inexplicable the behaviour is. 696 01:11:48,650 --> 01:11:55,700 And so updating this distribution or that longitudinal interaction is a great way to answer this gentleman's 697 01:11:55,700 --> 01:12:05,580 question as to what how do you learn that the robot model is that marriage is changing or the interaction? 698 01:12:05,580 --> 01:12:10,190 Okay. We do. We do still have like a minute. We have one other question that's come in. 699 01:12:10,190 --> 01:12:13,880 Could this kind of modelling be used for machine to machine interactions? 700 01:12:13,880 --> 01:12:18,890 If so, would you need to change the way you form models of what the other agent expects and have written? 701 01:12:18,890 --> 01:12:24,950 I'm really sorry if this is a stupid question. We know exactly is our life is spent dealing with stupid questions. 702 01:12:24,950 --> 01:12:31,730 That's fine. Actually, it's actually a great question and I thank you for asking this. 703 01:12:31,730 --> 01:12:37,280 Yes, it can be done, but no, I won't do it that way because machine and machine interaction, 704 01:12:37,280 --> 01:12:41,300 I would say that the thing that we are missing really and I hope Elon Musk won't 705 01:12:41,300 --> 01:12:45,770 be successful in changing this is we don't have a USP connected to our brains. 706 01:12:45,770 --> 01:12:54,680 So there is no brain to brain Grantland's. Okay, so you can only interact with these humans in this sort of just like we can't eat pizza or e-mail. 707 01:12:54,680 --> 01:13:02,720 We can't learn each other's mindsets by anything other than sort of watching their behaviour and making hypotheses. 708 01:13:02,720 --> 01:13:10,760 And so that's. What makes that human interaction more complex than human human interaction also has been more complex because of that. 709 01:13:10,760 --> 01:13:14,740 If I was designing robots, in fact, I would say they will have a hive mind. 710 01:13:14,740 --> 01:13:20,570 You know, if I have a whole bunch of robots that have to walk in the factory, there's no real reason why they have to deal with it. 711 01:13:20,570 --> 01:13:25,880 Am I being extra able to do that? I guess I'll have the same mind. I'll just make sure that you have the same mind. 712 01:13:25,880 --> 01:13:30,810 And then maybe somebody from North Korea will have a different hive mind. You just have to deal with that back. 713 01:13:30,810 --> 01:13:34,850 But there are cheaper solutions if it is machine to machine interaction. 714 01:13:34,850 --> 01:13:36,860 That's some pops up. This is all relevant. 715 01:13:36,860 --> 01:13:44,570 But there's a lot more that is really because of human in the loop here, because I still think that I'm confused. 716 01:13:44,570 --> 01:13:52,100 I'm not. This is the problem, right? What? The pandemic. We did everything on Zoome except eating pizza, which we had to do physically. 717 01:13:52,100 --> 01:13:56,720 And we are sort of inefficient that way. And same thing in terms of understanding each other. 718 01:13:56,720 --> 01:14:00,990 We have to do this in this inadequate way. So that makes things harder. 719 01:14:00,990 --> 01:14:06,110 But machine and machine, you can still do it in the hardware. But they are also easier ways because you can design the machine. 720 01:14:06,110 --> 01:14:09,770 And I'm hoping that we won't design that human ward. 721 01:14:09,770 --> 01:14:14,720 And I'm hoping that neural link doesn't become successful, at least not in my lifetime. 722 01:14:14,720 --> 01:14:19,220 I don't think I thing. I'm with you on that one. 723 01:14:19,220 --> 01:14:23,060 Thank you so much for a wonderful lecture. We very much appreciated it. 724 01:14:23,060 --> 01:14:27,290 And let me just say again, you are most welcome physically in Oxford when all this is over. 725 01:14:27,290 --> 01:14:32,270 We would love to have you. And I think you've you've now got a whole bunch of people who would love to meet you in person. 726 01:14:32,270 --> 01:14:37,520 So you'll be most welcome. Now, in the meantime, stay safe. And thank you again for those that haven't. 727 01:14:37,520 --> 01:14:42,800 I really urge you to look round pieces in the hill and also follow him on Twitter if you're 728 01:14:42,800 --> 01:14:48,590 interested in A.I. and for a humorous and informed commentary on the current state of A.I., 729 01:14:48,590 --> 01:14:52,370 there's no better place to look, I think. So thanks again, Raph, and take it. 730 01:14:52,370 --> 01:14:54,868 Thanks again for the invite. Mike.