Latest / Strachey Lectures / Strachey Lecture: The Quest for Truth in the Information Age

Transcript

Strachey Lecture: The Quest for Truth in the Information Age · Strachey Lectures

• 0:00
  So welcome, everyone, to today's strategy lecture.
• 0:04
  In case you don't know, the Strachey she lectures are a distinguished lecture series named for Christopher Strachey,
• 0:12
  a pioneer of computer science, known in particular for his contributions to programming languages and the notational semantics.
• 0:20
  And the founder of The Discipline at Oxford, the first professor of computing science in Oxford Street.
• 0:26
  He lectures a generously supported by Oxford Asset Management. The computer science department is very grateful for their support.
• 0:34
  Under normal circumstances, pre pandemic.
• 0:36
  This enabled us to hold lectures in the splendid venue and with very nice refreshments enjoyed by all the participants.
• 0:44
  We don't have those that is today. But what we do have is is an excellent lecturer.
• 0:50
  So it's my great pleasure to introduce today's straight lecturer, Professor Sonia Smedes.
• 0:56
  Sonia is Professor of Logic and Epistemology at the Institute for Logic, Language and Computation LLC at the University of Amsterdam.
• 1:06
  And since 2016, she has been the director of L.L.C., which is one of the major world centres for the study of logic and computation.
• 1:16
  Amongst her many activities and recognitions, too many to mention them all.
• 1:20
  But let me just mention that she was awarded the book of On the Human Prise for her work in Quantum Logic.
• 1:26
  She has held a Rosalind Franklin fellowship and she was elected to membership of the Academia Europea, the European Academy in 2019.
• 1:37
  Sonia's work has spanned a range of topics, notably dynamic quantum logic, dynamic epistemic logic, formal epistemology and social epistemology.
• 1:48
  Much of her recent work focuses on the use of logic to characterise and study information flow,
• 1:54
  opinion formation, social influence, informational cascades, homophily and polarisation in social networks.
• 2:02
  And the title of her lecture today is, I think, very resonant for us all.
• 2:07
  The Quest for Truth in the Information Age.
• 2:10
  Just before I hand over to Sonia, let me mention that if you have a question, please use the Q&A facility to send your question.
• 2:19
  You can find the Q&A panel by clicking on the icon with a question mark in it on the right hand side, top of your screen.
• 2:28
  After the lecture, we will have a Q&A session where I will read questions to Sonia and she will respond.
• 2:35
  OK, without further ado, I have a great pleasure in handing over to Sonia for the lecture.
• 2:40
  Sonia, please. OK, well, thank you so much for this invitation and thank you some some for a very nice introduction.
• 2:53
  It's a really great honour to give the stretchy lecture and computing science today.
• 2:59
  So the title is The Quest for Truth in the Information Age. So let me start by reflecting on worldwide connectivity as we know it today.
• 3:09
  So there are many advantages in worldwide connectivity, from finding information, from finding high school friends,
• 3:18
  creating new friends, finding jobs, finding supports, etc. So plenty of advantages, but with disadvantages.
• 3:27
  Also on the down side, there are many disadvantages as well.
• 3:30
  So we see the spread of misinformation. We see online believe manipulation, online radicalisation, many cases of online bullying in schools,
• 3:44
  the formation of informational cascade's and echo chambers to which we come back later on.
• 3:50
  So that brings me to the question can truth survive the information age?
• 3:55
  Now, answering this question requires a study of the mechanisms for mass information
• 4:02
  aggregation and information exchange in relation to our ability to track the truth.
• 4:09
  So I will focus on the social and epistemic benefits and costs of such mechanisms,
• 4:16
  and I will reflect upon the successes and failures of individual versus collective rationality.
• 4:24
  So the approach that I adopt brings together many insights from a series of different fields.
• 4:31
  So one is I as a magician, I use logic to model the dynamics of information and we look specifically at knowledge and belief.
• 4:43
  We take input from belief, revision, theory and philosophy and another area from philosophy.
• 4:49
  We use formal epistemology and a particular social epistemology.
• 4:54
  Epistemology is a theory of knowledge.
• 4:57
  We use input also from judgement, aggregation and social choice theory, and our latest work uses more input from network theory.
• 5:07
  So let's focus on this aspect of tracking the truth and pose the first question.
• 5:14
  So how the group successes and failures to track the truth?
• 5:18
  How are they related to individual successes and failures to track the truth?
• 5:23
  Not answering this means that we have to look at what is known as wisdom versus madness of crowds,
• 5:31
  and maybe more specifically, what actually does a group know?
• 5:35
  So what is the knowledge that a group can possess? So I'll call this the group's potential knowledge.
• 5:41
  And what I mean by that is the knowledge that the members of a group may come to possess
• 5:48
  by combining their individual knowledge and using their joint epistemic capability.
• 5:55
  But in what sense can a group process beliefs or knowledge?
• 5:59
  Of course, you can infer this by looking at the group's behaviour.
• 6:03
  This is the joint actions that they can do, that the members of the group can perform after combining their individual knowledge.
• 6:13
  So that still brings the question, so how smart is a group?
• 6:19
  So does a group no more than each of the members that that is in that group.
• 6:25
  So on the one hand, we can say yes, yes, because the new information that was initially unknown to any of the agents may be combined,
• 6:38
  may be obtained by combining individual different pieces of private information possessed by the different agents.
• 6:46
  So in potentially we know more as a group than each of us individually.
• 6:50
  But the opposite may also happen when some, maybe all individual knowledge is made inaccessible to the group.
• 6:59
  And now you can think, well, maybe at first sight the opposite never happens.
• 7:04
  Maybe the group's potential knowledge includes everything known by any of the members.
• 7:10
  So that exactly is what is called wisdom of the crowds.
• 7:14
  And there's a lot of popularising literature on that also book with this title.
• 7:20
  So the wisdom of the crowds is this idea that the potential knowledge of the group is
• 7:24
  typically much higher than even the knowledge of the most expert member of the group.
• 7:30
  Now, if you looking into this a little bit more detail, you see that there are different examples of wisdom of the crowds.
• 7:38
  And I will make a distinction between two kinds of examples. And each of those is based on.
• 7:44
  Different forms of information aggregation, and they have different explanations.
• 7:50
  So if we focus on the first type, then you would say, OK, a different way of aggregating information would be by communication.
• 8:02
  So a group can actualise a piece of group knowledge by communication on the one hand.
• 8:08
  And on the other hand, it would use a different method for aggregation, which would be pulling the opinions together, some form of judgement,
• 8:17
  aggregation, and how good a group will be in realising its epistemic potential or bringing about the potential group.
• 8:26
  Knowledge will depend on each of these methods, but also on different factors.
• 8:31
  So the different factors that play a role are well, first of all,
• 8:36
  it's important what the topology of the communication network will be like, who will talk to whom, but also trust plays a role.
• 8:45
  So the mutual trust graph is important.
• 8:48
  So how trustworthy do you think that the information is that you get from another agent or a subgroup of agents?
• 8:56
  And self trust will play a role, so this includes the threshold of what is needed for an agent to change her beliefs.
• 9:05
  So from which point onwards, can you be influenced or are too influential to be subjected to information that comes from others?
• 9:15
  And the last point here on that slide is about the the agents epistemic interest plays a role.
• 9:23
  This means her own agenda, her own questions, what she is interested in and how she filters the information that she gets her learning goals.
• 9:35
  So if we look a little bit more closer, then you see that these two different ways of obtaining or pulling information together,
• 9:44
  one via communication and the other via some aggregation method, they lead to two different notions of group knowledge.
• 9:52
  So the first one, the first notion of group knowledge that we see would be the one that is based on communication.
• 10:00
  So so here you see that the agents will have a piece of information or a piece
• 10:09
  of evidence that typically is truthful and the agents share their knowledge.
• 10:14
  And examples of this type of group knowledge would be joint authorship on a paper.
• 10:21
  So agents that collaborate on the project, such as maybe a big science project like the Human Genome Project,
• 10:28
  maybe proof of a mathematical theorem for us last year, but also a deliberation that happens in the hiring committee.
• 10:38
  So these are all explanations were individual experts.
• 10:43
  They collaborate together, they share their knowledge and they communicate doing so.
• 10:48
  And that brings evolved form of group knowledge. And the form that it brings about is typically described as.
• 10:55
  So the explanation in epistemic logic and epistemology that is given for this is distributed knowledge.
• 11:02
  So distributed knowledge is a specific form that is based on information sharing via communication between agents were as an example.
• 11:11
  I know p you know, that implies. Q So we pull the information together and we use our reasoning powers.
• 11:19
  So we do multiple opponents in this case. This means we both know P and by modest performance, we both know Q.
• 11:27
  So that was the first type of knowledge. Now the second type is an entirely different type.
• 11:32
  So here, in contrast to the first type, communication is not allowed.
• 11:38
  So here we see agents that aggregate independent opinions so they aggregate their beliefs.
• 11:45
  And these beliefs are based on private observations of pieces of evidence that they have.
• 11:52
  But in this case, the evidence is soft. So it's possibly that the evidence is fallible and it is important here.
• 12:01
  This is a different type of group knowledge because it requires agents to pull together their information based on independent observations.
• 12:10
  So the evidence that they have, they make an independent observation of this and they base their opinions on that.
• 12:18
  So examples that you see in the literature are they include the verification of experimental results.
• 12:25
  And in the popular science literature, you will see examples that refer to there is an important example,
• 12:33
  the case described by Francis Galton in nineteen of six, he went to a livestock fair or that was an ox on display.
• 12:42
  And what Francis did was he collected all the so there was a contest where people could win a prise
• 12:50
  if they guessed the weight of the ox correctly and they each had to ride their guests on the ballot.
• 12:56
  So Francis collected all these ballots. And what he saw was that in the end, if you would take the mean of all the guesses,
• 13:04
  take the average, then that actually gets to the almost exact weight of the ox.
• 13:12
  So he saw that collecting that independent information from all the agents and then saying this gets you to almost a perfect answer.
• 13:22
  Other examples in the literature are, for instance, counting jellybeans in a jar.
• 13:28
  This is also an example that was done by Jack Trainer already in the 1920s.
• 13:34
  In his classroom, he took a jelly bean jar to his classroom, asked the students to to give an estimate of the number of beans in the jar.
• 13:43
  And he saw that if you take the mean of all the gases, you get almost the correct answer.
• 13:50
  So here indeed, this works very well. This method for aggregating the beliefs works very well.
• 13:56
  But it is based on this idea that no communication is allowed and the guesses of the individual participants are independent.
• 14:07
  Now, we can illustrate that with a more concrete example.
• 14:13
  So this will be what I call the first or an example. So let's take a sequence of individual agents.
• 14:20
  And I enumerate these agents a want to. And and the idea is that these agents entered into a room.
• 14:27
  So they stand in a line. So it's a sequence. They enter into that room.
• 14:31
  And it is common knowledge that in that room there is one of two possible urns they don't know and they can't see which one from the outside.
• 14:39
  So it is either urn w an urn w contains two white walls, one black wool or A or B and urn B contains two black males and one white ball.
• 14:53
  So that is common knowledge that one of these answers in the room and in person goes into the room there,
• 15:00
  a ball from the urn looks at it and makes a guess about which urn is placed in the room.
• 15:06
  So the guess is either W or B. And in this case, the guesses are secret.
• 15:11
  So no communication is allowed. So imagine the launch sequence of agents, each entering and having a guest that may be writing that guess on a ballot.
• 15:23
  Now, at the end, the poll is taken of all the people's guesses.
• 15:28
  And what we see is that the majority guests actually will then be what we call the potential group knowledge.
• 15:35
  Given the size of the group gets to infinity, the group gets virtual knowledge of the real estate,
• 15:42
  with a probability approaching to one of the size of the group is large enough then the truth.
• 15:48
  Tracking power of this group via this method is excellent.
• 15:53
  And the standard explanation for this can be given by some variation of council says here, which is essentially based on the law of large numbers.
• 16:04
  And what it says here is that when you perform many independent observations, the individual errors are the pieces of private evidence.
• 16:13
  Supporting the false hypothesis will be outnumbered by the truth supporting evidence.
• 16:20
  So have statement of the theorem a little bit more specifically correct state.
• 16:27
  So the term Sesto, the probability that the majority selects the correct alternative approaches to one when the size
• 16:35
  of the group and goes to infinity whenever the following three constraints of conditions are met,
• 16:42
  namely, one, that each individual makes an independent decision based only on private information,
• 16:51
  the likelihood of the private signal to be correct has to be bigger than than a half.
• 16:59
  And in this case, the choice has to be binary. So we have two mutually exclusive alternatives that people vote on or that they choose from.
• 17:09
  And that's an important term.
• 17:11
  It has also been used in a lot of different contexts, so the term of Canossa has been used to argue in favour of epistemic democracy.
• 17:22
  That is the idea that in normal circumstances, the democratic forms of belief aggregation are indeed conducive to the truth.
• 17:31
  So I refer you here to the work of Christian listing Goulden for this and what they did in their work.
• 17:40
  I've quoted here, I've put the reference for the paper is that they actually extended the term in their argument to go beyond the binary choice.
• 17:49
  That was the last condition in the in the statement of the term.
• 17:53
  And so they looked at voting scenarios that involved many correct options and they still saw that that mattered.
• 18:03
  So this extended version of the concept here is very good.
• 18:07
  And as a as a belief aggregation method that is conducive to the truth.
• 18:16
  Now, what if the jury tarim, if we look at it?
• 18:22
  Of course, it assumes that the agent's opinions are based only on independent observations, like when the people entered that room in the urn example,
• 18:30
  they look at the bull and they have to make their guess independently and they have their own private independent evidence for doing so.
• 18:40
  Now, what happens if we allow some communication of observations?
• 18:44
  Right in the example to now, communication was not allowed.
• 18:48
  If we do that, if we allow communication, then the opinions will no longer be independent, which means that the Condorcet jury theorem will not apply.
• 18:59
  And intuitively, we might think that after communication,
• 19:03
  the agents entering the room in their own example later on now they would have maybe more information than just their private observations.
• 19:11
  So you would intuitively expect that the group's knowledge would increase.
• 19:17
  However, as we'll soon see when I extend the example, this is actually not what happens.
• 19:24
  So that brings me to the fragility of group knowledge.
• 19:27
  So we've talked about two forms of wisdom of the crowds and two specific forms that we're linked to, two types of group knowledge.
• 19:37
  So both types are actually very fragile because any change in the conditions that are necessary
• 19:45
  to bring about or actualise that form of group knowledge can lead to under optimal results.
• 19:53
  So let's look at the first type of group knowledge, which is the one that is based on communication.
• 19:59
  Now, in this case, if the information exchange is limited by the network or by the degree of trust or by the agents agendas,
• 20:09
  then the group will only actualise something much less than full distributed knowledge.
• 20:17
  So this type of group knowledge can fail. And the reason for the failure is that we would say that it this some kind of selective hearing.
• 20:26
  So in this case, the agents listen only to certain sources, mainly only their friends.
• 20:37
  They listen only if it fits the information, fits their own interests or fits their own agenda.
• 20:45
  And maybe they only accept new information if it is consistent with their own
• 20:49
  information that they already had from before or their pre-existing beliefs.
• 20:56
  So in my work with my co-authors,
• 21:00
  we looked at the specific example which we called the curse of the committee because of many deliberating committees.
• 21:08
  After long sessions of deliberations, information sharing, a lot of argumentation,
• 21:14
  it often happens that the only thing at the end that the committee agrees on
• 21:19
  is actually what they already agreed on before the start of the deliberation.
• 21:24
  Of course, in these cases, the potential knowledge of the group seems to boil down to the groups prior common knowledge.
• 21:32
  So the information that was common knowledge at the start of the deliberation,
• 21:36
  but not at the end, and that is typically much poorer than distributed knowledge.
• 21:41
  Distributed knowledge is the ideal of what they could reach. So in a sense, this is the worst possible outcome.
• 21:49
  And now one reason why this happens is the selective hearing principal were an agent's actual or potential
• 21:59
  knowledge is limited by her own epistemic interests and her own questions and the things that she's interested in.
• 22:07
  So the liberating agents here are process only the information that is relevant to their own questions or issues.
• 22:15
  So in joint work with example Baltar and Rachel body,
• 22:19
  we explain this example of the curse of the committee by looking at the at each of the agents epistemic
• 22:26
  interests that they have to actually give a full formalisation of this which are not doing this talk,
• 22:33
  but I will give a reference to it. So we use a version of interrogative epistemic logic that is an epistemic logic with an interrogator component,
• 22:44
  which means we are also modelling the questions on the issues that the agents have,
• 22:49
  and we put a specific constraint on how we represent the knowledge of the agents,
• 22:55
  including the fact that one can only know things that are actually relevant to the issues
• 23:00
  that other agent has so relevant to the questions or the agenda of that specific agent.
• 23:07
  And the same constraint also applies on the dynamics that we that we model.
• 23:12
  So so the dynamics means that here we look at public announcements and that has to comply with the selective hearing principle.
• 23:24
  So an agent only learns from an announcement if it's relevant and only learns from an
• 23:32
  announcement what the relevant consequences are relevant to the agent's own agenda.
• 23:38
  When you do that, then it's easy to see that this indeed can lead to radical disagreement.
• 23:45
  Now in the paper rules to formalise what the actual group knowledge is in such an agenda driven context,
• 23:56
  so what we call the the groups, the group Jews epistemic potential or group potential group knowledge,
• 24:05
  we the this case superscripts that will amount to whatever will become common knowledge after each
• 24:13
  agent in the group updates her knowledge with the group's information that answers her questions.
• 24:22
  So this potential will typically lie in between initial common knowledge and what is known in the literature as distributed knowledge.
• 24:32
  So what we saw in our paper is that distributed knowledge and common knowledge,
• 24:38
  which are the main concepts from epistemic logic and the standard literature,
• 24:44
  they can actually be seen as too extreme versions of this potential group knowledge,
• 24:49
  giving us the upper and the lower limits of the group's epistemic potential.
• 24:55
  So if every agent in the group is interested in all the information processed by all everybody else, so when the agendas are very cohesive,
• 25:04
  then the potential norm, which is maximal and then the potential group knowledge will be equally distributed knowledge.
• 25:12
  But on the other side, if the agents have very dissimilar agendas, namely the questions they are interested in are orthogonal to each other.
• 25:21
  So no agent has opened questions that can actually be answered by the group than a potential group.
• 25:28
  Knowledge is very minimal.
• 25:29
  And in that case, as in the curse of the committee example, they will fall back on what was already common knowledge of START.
• 25:38
  So the lesson that we learn from this is that a group formed of agents with very dissimilar agendas will have a very poor epistemic power,
• 25:49
  they say, actually with what they already knew. And when you have a committee or a group with a very cohesive agenda, with dissimilar expertise,
• 26:00
  still need specific experts coming together, they will have a very strong epistemic potential.
• 26:06
  They can actually learn from each other.
• 26:08
  And here I gave also the reference to the paper in which we explore these results, where we make this formally precise.
• 26:18
  So the curse of the comment,
• 26:19
  the example that we investigate that can actually be related to what is known as the common knowledge of fact and social psychology.
• 26:29
  So that common knowledge of fact refers to the case where groups tend to focus more
• 26:35
  on information that is already common rather than on sharing their private evidence.
• 26:41
  Now, the specific relation to this common knowledge of facts should still be explored because
• 26:47
  it's my understanding that they actually talk about another notion of knowledge,
• 26:53
  which is not really common, but more something like mutual knowledge of what everybody knows.
• 26:59
  But still, it's interesting that this effect has been has been studied and points out that information that the committee had already before the
• 27:07
  deliberations starts is something that the group tends to focus more on that rather than on sharing their own private information.
• 27:17
  So if I go back now to our information society, then what we see happening is that.
• 27:24
  People choose with whom to associate because information technology also enhances visibility for more choice based on our preferences or agendas,
• 27:36
  and having more choice is a good thing. At least that's what we think.
• 27:41
  But these examples also show that information exchange can be filtered through these choices.
• 27:49
  So more connectivity without more epistemic sophistication can actually accelerate effects of polarisation.
• 27:58
  The formation of Cascades echo chambers. So these examples bring me to the to the fragility of group knowledge again.
• 28:09
  But now let's look at the second type. So we've we have already seen the fragility of group knowledge that is based on communication and
• 28:17
  what can go wrong when people filter incoming information through their agendas and interests.
• 28:23
  But let's look at the second type of group knowledge, which is not based on communication,
• 28:28
  but which is based on some judgement aggregation method of voting out where people have to bring together independent opinions.
• 28:39
  So this type is also very fragile and also prone to failure because in this case, any breach of the agent's independence.
• 28:47
  So any communication can lead the group astray.
• 28:52
  And typical examples here are indeed the formation of an informational cascade, which you can think of as some form of rational epistemic bandwagon's.
• 29:03
  Forms of groupthink and notion of pluralistic ignorance is the scenario where we point to a lack of common knowledge leading to group.
• 29:15
  The group suppresses generally shared results. And another example would be radical disagreement, which will lead to group polarisation.
• 29:26
  So let's zoom in on the first example here on the formation of an informational cascade and on its social and epistemic features.
• 29:36
  So so imagine a sequence of people.
• 29:40
  So the line of of agents as an example and the informational cascade happens when the people in that line base their
• 29:54
  decisions by observing the decisions of the predecessor and they follow their the decisions of their predecessors,
• 30:02
  ignoring their own private evidence that they have. So these situations happen.
• 30:09
  And it is not necessarily the case that this is just mindless imitation that we're imitating our predecessors.
• 30:16
  Now, these people are not brainwashed. They're not part of mass hysteria.
• 30:20
  So actually, this happens in cases where people are assumed to be rational and they have a rational choice.
• 30:29
  So let's go back to our own example, which is very fit to study this,
• 30:33
  so take the same or an example that I have had before where the standard explanation was given by the commerce security here,
• 30:42
  except now I will allow a limited form of communication.
• 30:47
  So when each agent makes her guess for W.O. or be the guest now is immediately publicly announced.
• 30:55
  So they go into the room, the drawable they look at it and they make a guess.
• 31:00
  But instead of writing it on a ballot, they go out and they make a public announcement.
• 31:05
  And what we see is that an informational cascade arises with a probability of at least one of the nine.
• 31:13
  If the first two votes that are drawn are of the minority type.
• 31:18
  So suppose that the person in the room is or w and the first two agents who enter the room, they take a ball and it is a black ball.
• 31:27
  Then the first two agents rationally will guess B. And they make their guesses publicly.
• 31:35
  There is a public announcement of their guesses to everybody else, everybody else can hear it,
• 31:39
  then all the agents will rationally follow and they will all select the wrong option B. So the wrong option will then be unanimously selected.
• 31:50
  So now we can ask the question, is this rational? And according to an analysis, actually, the answer is yes,
• 31:58
  because given the available information that the agents have been agents that follow probability
• 32:06
  theory and based based update's rule when they are interested in individual truth tracking,
• 32:12
  they will behave exactly in this way. So this means that individual Bayesian rationality can lead the whole group to the wrong answer.
• 32:22
  So to group irrationality, some people throw doubt on the Bayesian proof because it doesn't make explicit the higher order reasoning of the agents.
• 32:36
  So you can do the whole analysis of what happens in the unexampled right down a right of the proof.
• 32:43
  And what you will see is that it has credence is there for a simple belief,
• 32:48
  but it doesn't take into account the high order reasoning in higher order beliefs of the agents.
• 32:53
  So the question we ask ourselves is that if agents who who have a theory of the
• 33:01
  mind and who can reflect on the overall protocol that they can reflect on,
• 33:06
  on the whole scenario that is happening and they can realise that they are part of a cascade and they reason about other agents minds.
• 33:16
  Can they avoid the cascade so they can piece everything together?
• 33:20
  They have a full theory of the mind. Can they avoid the cascade?
• 33:24
  And that can happen for a number of cases. The answer may be yes, but it is not for an example.
• 33:31
  So whatever reasoning powers that we give our agents,
• 33:35
  it's not going to help them to make a rational decision to break this cascade of what we did in our paper.
• 33:43
  So I've given the reference to the paper here. Below the slides is that we we have a proof for that.
• 33:50
  So we use dynamic epistemic logic to prove three, prove the argument.
• 33:56
  But now in an epistemic logic setting where you can make the reasoning powers of the agents explicit so the agents have unlimited reflective power,
• 34:06
  these are super duper agents. And I mean the tools that we have for this, they work very well because indeed,
• 34:13
  epistemic logic incorporates all the levels of knowledge and belief that agents have.
• 34:19
  And dynamic epistemic logic allows us to add the knowledge and belief that agents have about what is going on,
• 34:27
  about the protocol, about the informational events. And the conclusion that we have is exactly the same conclusion as you would get in the case.
• 34:36
  But now that the full logical omniscience, higher order reasoning,
• 34:41
  reflecting that you're part of an informational cascade is not enough for you to rationally make a different decision and get out of this cascade.
• 34:52
  If you look at the informational cascade, note that the the strength of evidence that agents have doesn't get higher,
• 35:01
  doesn't rise as the cascade progresses, because the third guess in this example itself is actually uninformative.
• 35:11
  Only the first two guesses, if you think about example, are informative.
• 35:15
  So the information of the fourth person or the person on the line is actually exactly the same.
• 35:23
  Of course, a reality we see that people are reinforced when they see millions of other agents and from them all making the same decision,
• 35:33
  even though maybe the only informative part are the decisions made by the by the first two agents and not everybody else around it.
• 35:42
  But it might increase. In reality, we see that people tend to have a stronger level of belief,
• 35:48
  even though that is done in that case for the unexampled, that would not be justified.
• 35:54
  What we did in our work is, well, it's a different paper here on the dynamic epistemic logics of diffusion and prediction and social
• 36:02
  networks is that we generalise the case of Cascade's in in this linear networks to arbitrary networks.
• 36:11
  So arbitrary, I mean, any type of topology, not necessarily linear line.
• 36:16
  So we looked at how fashion and behaviour propagates in a network of agents with its theory of the mind.
• 36:25
  And in particular here, these agents can anticipate or predict the behaviour of this of their friends.
• 36:33
  So they use their higher order,
• 36:35
  reasoning power to get reason about what the other agents are going to do and they can predict the behaviour of the others.
• 36:44
  So what we saw the conclusion of that paper is that knowledge does make a difference.
• 36:51
  So if you look at the speed of how fast?
• 36:53
  It emerges then the epistemic ingredient is very important because knowledge can speed up the spread of this cascade formation,
• 37:03
  so the spread of behaviour while on the other hand, uncertainty can slow it down.
• 37:10
  Even more, if it is common knowledge that everybody knows everybody's preferences or behaviour,
• 37:17
  and of course, also they need to have some knowledge about the network.
• 37:21
  So the friendship relation. So who's connected to who then, those gates?
• 37:27
  Could the principle be reached immediately? So that will immediately speed up the formation of the of the whole cascade.
• 37:34
  So, again, the analysis that we did for that paper is fully based on dynamic epistemic logic where you can make
• 37:42
  these aspects also the prediction of behaviour of others fully explicit in the logical language.
• 37:51
  Now, these cascades happened in reality, and there are many examples of newspapers are full of them.
• 37:58
  But let me give you one example that comes from the Netherlands.
• 38:03
  So this is an example that hit the newspapers and it was called Project X.
• 38:09
  So it brings me to 2012. So we we are in the city of Hama and Hanna is close to Thorning in the north northeast of the Netherlands.
• 38:21
  So in 2012, Miftah invited all her Facebook friends for her 16th birthday party and she made the
• 38:31
  message public on Facebook so that friends could bring in principle other friends.
• 38:38
  And what we saw is that one friend invited five hundred other people, so five hundred other friends.
• 38:45
  And then a day later, sixteen thousand people signed up for the party.
• 38:49
  So by now, Metho and her parents, they were getting afraid of this and said, like, OK, this is getting out of hand.
• 38:57
  This was not the intention. We are going to cancel the party. The party was cancelled.
• 39:02
  But what happened on Facebook is that other organisers took over.
• 39:06
  They called it Project X and they they kept sending out the announcement for the party.
• 39:12
  So very soon. Thirty thousand people signed up.
• 39:15
  And exactly on her birthday, four hundred thousand messages were posted on Twitter.
• 39:21
  And in the evening for the birthday party, still thousands came to harden.
• 39:26
  And this and the very bad side ended with riots with the police metho and her friends and her family.
• 39:33
  They had to leave the town. They couldn't. The whole street, I think, was evacuated, or at least her family had to evacuate.
• 39:44
  Examples of this type and saw, especially via Facebook and Twitter, we can see these online cascade's emerging.
• 39:54
  If you look back at the classic example, specifically in the iPhone example,
• 39:58
  then we saw that the cascade happened because agents were rewarded for their individual through tracking capabilities.
• 40:08
  So they were I mean, we didn't make the reward or the payoff explicit in their own example.
• 40:13
  But the way to set up is they actually only looked at their individual evidence and they tried to get each of them individually to to make it correct.
• 40:25
  Guess for the earth to be W or B now, you could change that.
• 40:30
  You could make the playoffs explicit and change these payoffs, rewarding the agents,
• 40:36
  not when they got to the individual rights of their individual guests would be correct,
• 40:43
  but maybe rewarding the agents only if and only if the whole majority, the whole group tracked the truth.
• 40:51
  And in that case, the cascade should not form because rational players will then disregard the information that
• 41:00
  they received themselves and they'll take the information that they got themselves into account.
• 41:05
  But disregard the guesses, the information that we received from the others, simply guessing the urn that matches to the colour that they saw.
• 41:15
  And they should take full advantage of the recessed herem, which was perfect,
• 41:20
  the tracking the truth when they disregard the communication of the others.
• 41:27
  Now, that means that if we look closer at the role of communication in this, so.
• 41:35
  In the cascading earn example, agents were allowed to communicate and then the Cascades could form,
• 41:42
  but it happened because there they were allowed to communicate their opinions,
• 41:48
  but not their justification, not their private evidence or their private reasons for making a certain guess.
• 41:57
  So communicate communities who do communicate all their private evidence,
• 42:03
  you could think while this solves maybe the problem in the informational and the unexampled,
• 42:09
  where we see that informational cascade happening because of people in that case would announce their private evidence again, should not happen.
• 42:18
  But that is not always the case.
• 42:20
  So communities who do communicate the private evidence aren't necessarily always guaranteed to track the truth and to do better.
• 42:30
  So there are many other cascades that can happen where to enrich too much communication
• 42:37
  actually is counterproductive and where too much communication leads us to a problem.
• 42:45
  So this is in formal epistemology and in social epistemology.
• 42:51
  This is what is called now the Solman effect. So it refers to Kevin Zillman, who works at the University of Pittsburgh,
• 42:59
  and he studied the effect of connectivity on the ability of the scientific community to track the truth,
• 43:08
  because we as scientists, we would think that more common connectivity in science is better.
• 43:13
  Right. So lots of conferences, lots of communication. That should be good.
• 43:17
  Now, what Kevin found and that shocked the community is that while the answer to this is actually no,
• 43:25
  there are contexts in which a community of scientists as a whole will be more reliable
• 43:33
  when the members are actually less aware of the experimental results of their colleagues.
• 43:39
  And the reason for that is that too early elimination of correct alternative theories or options or experiments that can actually block the
• 43:51
  group from tracking the truth so they influence each other and they might just decide not to explore these alternative options anymore,
• 44:00
  which might be the key for the whole community to get it right.
• 44:05
  So let me give a quote of Solomon from his 2007 paper.
• 44:10
  So he says that in circumstances where speed is very important,
• 44:15
  but where we think that our initial estimates are like likely very close to the truth, then connected groups of scientists will be more reliable.
• 44:25
  So if we already happened to be close to the truth, that will do well. On the other hand, where we want accuracy above all else,
• 44:33
  then we should prefer communities made up of more isolated individuals so that the number of conferences that we see exploding in our communities,
• 44:46
  maybe not necessarily always good for the whole community to get to the truth.
• 44:54
  An example of a cascade in which more communication is harmful,
• 44:59
  all the while there are many examples, but it would lead, for instance, to believe polarisation.
• 45:06
  So believe polarisation is an informational cascade that is then produced by the widespread mutual communication of a preliminary test results.
• 45:18
  So not the final results. And that leads to two early elimination of alternative options or theories that are out there.
• 45:27
  And together it leads to a strengthening of a biased majority opinion, and that biased majority opinion will not track the truth.
• 45:38
  So in these scenarios, more communication means that it is much more likely that polarisation can actually occur.
• 45:49
  So then the question is, how can we actually avoid this problem?
• 45:53
  So how can we get out of that? Well, Spirit Sparks Networks would be would be an answer one would need to preserve the opinions
• 46:04
  independence for long enough so that enough alternative theories and approaches are checked out.
• 46:12
  As a result, the probability of a cascade will then be lower.
• 46:16
  So on these past networks, they are more likely to be through tracking in the long run.
• 46:24
  Now, of course, there is a Trade-Off between efficiency and speed and on the one hand and the reliability on the other hand,
• 46:35
  by reliability I mean here the avoiding the avoidance of a cascading, merging or polarisation happening.
• 46:42
  So because these past networks there may be very good for science,
• 46:49
  but there are also very slow at converging to any common opinion whether that opinion is true or not.
• 46:55
  So a very connected network converges very fast.
• 47:00
  And so when it hits the truth, it does so in a very efficient way and much more efficient than a sports network.
• 47:09
  But we are, of course, not guaranteed that it will hit actually truth if it is very connected.
• 47:14
  That's exactly what Kevin shows in his work.
• 47:19
  So. So indeed, we have this trade off between speed and reliability and what the community would meet and also what the community of scientists would
• 47:28
  meet would be a communication network that lies somewhere in the middle ground between extreme sparsity and extreme connectivity.
• 47:40
  Now, of course, there are agencies out there that know these theories and they can use it against the scientists.
• 47:49
  So you can if you know that alternative theories are necessary to progress science and that
• 47:58
  more alternative theories need to be investigated before actually full opinions are formed,
• 48:05
  then that can be used to create doubt and to undermine scientific truths, maybe when they have already been established.
• 48:14
  So when a community is too slow in tracking and maybe even just to convey the truth,
• 48:20
  even when they know it already, it will leave room for unwanted effects.
• 48:25
  So so I recently read an article in The New York Times and the headline there of that
• 48:32
  article was Influencers say that they were urged to criticise the fire's SURFAXIN.
• 48:39
  So disinformation efforts to reduce public confidence and covid-19 vaccinates try to role social media commentators in France and Germany.
• 48:49
  So you see that this idea that diversity of attempts and theories is good for the progress of science.
• 48:59
  Less activity is good. We don't need to converge to the truth too fast can also be used against us.
• 49:08
  So, for instance, when people build an on their mind, a truth that has been already established and tried to communicate that very fast.
• 49:18
  So if you take stock now of the different epistemic and social dilemmas that I've touched upon,
• 49:24
  then a main factor that you'll see is that indeed rational individuals can form
• 49:33
  irrational groups and rational groups might well be composed of irrational individuals.
• 49:40
  This, of course, is nothing new in formal epistemology. This is known as the independence thesis.
• 49:48
  And this refers to a paper that was written and published in Philosophy of Science in 2011 by Michael Wilson Commensal and they could Dang's.
• 49:59
  So they also on their paper, give lots of examples that come from philosophy, science game theory, social choice theory.
• 50:10
  Now, the question is, what kind of what can logic do?
• 50:14
  So can logic help us understand these social epistemic dilemmas?
• 50:21
  And I would argue that, well, yes, logic can provide us with good models to reason and predict about this epistemic distortion.
• 50:31
  So for for a diagnosis that we want to to put on the problem, then, yes, logic is a very good tool.
• 50:39
  And examples are given by the examples that I gave in the presentation here.
• 50:44
  But so we use dynamic epistemic logic to investigate the formation of these informational cascades.
• 50:52
  We also study a and social properties in interaction when we analyse the the behaviour of both online behaviour of influencers.
• 51:05
  What triggers the formation of a polarised network of these filtering mechanisms that you might see in an echo chambers,
• 51:13
  the effects of homophily, etc.? So there is, of course, network theory out there and a lot of simulations and models are provided by network theory.
• 51:24
  But most of the models on the market don't take the epistemic ingredients into account.
• 51:30
  And I think for studying groups of rational agents, the study criteria are crucial.
• 51:37
  So that leaves another question. So can computer science help us verify and debunk our social software, diagnose the weaknesses,
• 51:47
  provide tools that could be used to prevent the worst forms of groupthink?
• 51:53
  And maybe that's a question for computer scientists and the audience.
• 51:58
  So that brings me to my last slide. So here's my conclusion.
• 52:05
  So let's look back at the question of the lecture. So will the facts.
• 52:10
  And then, I mean, the actual facts, things that are true in the world eventually emerge from the noisy network of irrelevant data,
• 52:17
  self confirming rumours, idiosyncratic likes, or, to put it more bluntly, will truth survive the information age?
• 52:25
  Now, I am an optimist. So according to me, knowing the mechanisms for mass information aggregation, for information exchange,
• 52:35
  we have to learn and know the epistemic benefits and costs and all that is just a start, but it is a very good start.
• 52:43
  It will hopefully allow us to perform some of the most urgent interventions very soon.
• 52:50
  I know that some interventions already are being taken.
• 52:53
  So there is also when you look at how the platform, the main social platforms are being organised, there is input from the legal side,
• 53:02
  their actions on decisions made by politicians, the platforms themselves, also how fact checking algorithms, tool to fight fake news, etc.
• 53:11
  So that is very good. But I also think that this epistemic factors need to be studied more and they play a crucial role in the formation of these
• 53:22
  online cascade's and all these negative side effects and distortions that we saw for the different types of wisdom of the crowds.
• 53:32
  So with this, I think this concludes my lecture.
• 53:36
  So thanks for listening and for attending, and I don't give the floor back to some.
• 53:44
  Thank you very much. But that was a wonderful lecture. And you showed us how logic and fútbol methods can reach deep into a lot of the prominent
• 53:54
  features of our information age in a way that I think many of us will be really illuminating.
• 54:02
  So, OK, so we we couldn't go to Q&A.
• 54:10
  We if you do have questions for Sonia, please send the questions on the on the Q&A panel, which you can find, as I said, with the icon.
• 54:23
  But the question mark. So let me just start with a couple that have been sent, one coming, I think, from a philosophical side,
• 54:33
  simply asking, since you speak of truth, what is the notion of truth that you're working with?
• 54:39
  Is it as simple as a simple bubble theory account, enough to fully encapsulate truth?
• 54:47
  So that's the question. Oh, yes. The notion of truth.
• 54:51
  I mean, it's the one that we use and logics of correspondence to the facts, I would say.
• 54:57
  So it's a very simple one. It's not so that many different notions of truth out there.
• 55:02
  So, for instance, coherence, theories of truth, where information needs to be coherent or things like that.
• 55:07
  So that's not the one I'm referring to, but just something that corresponds to basic facts and reality.
• 55:12
  So something that can always be checked. So, yes, of course, also on the logic side there,
• 55:23
  there are logics to deal with different theories of truth and to formalise that explicitly, but that we don't do so.
• 55:29
  It's always mixture. An external factor to the models that we have is somewhere a little point that says this is the true option.
• 55:39
  Yes, this is the model who has to build in that option and he needs to have that information available.
• 55:46
  So we assume that that is possible to get that information available, even if the agents don't know it.
• 55:52
  So agencies don't necessarily have to know that something is true.
• 55:56
  Knowledge is truthful, but not necessarily all the agents in the model will have access to that point,
• 56:02
  that is the real world, but the modellers should have that have access to that point.
• 56:07
  No. OK, thanks. So another question, what would your what would your suggestions be for us to come?
• 56:19
  I mean, this is the sort of the big practical question that, you know, they so the question is,
• 56:25
  what would your suggestions be for us to combat fake news in today's super connected society with all the powerful social media and,
• 56:38
  for example, Facebook, whose whole kind of mission is to make people more connected?
• 56:44
  So that's what they're all about. Oh, yeah.
• 56:48
  So there there are many things that you can do to combat fake news.
• 56:51
  So I think an important one is to educate people, self critical thinking, self reflection, sending our students to school,
• 57:02
  giving them logic classes from the beginning, that that is crucial for critical thinking because and I do think that will help.
• 57:10
  So that is one thing.
• 57:11
  So educate people so that they can they can think first when they see something, without retweeting it, without sending it further,
• 57:22
  if it is not coming from a reliable source or not from something that they're absolutely certain about.
• 57:30
  So that would be one thing then.
• 57:33
  Currently, I think some universities designed this fact checking games that people can play so people can then see how well they how good they are.
• 57:44
  In fact, checking some of the fake news themselves and can they discern the fake from the real news articles.
• 57:50
  So these games are excellent because means that people will then think about these things and they can they can learn.
• 57:56
  And are we also see that the younger generation actually is already much better at doing
• 58:02
  so than the ones that are growing up with all the social media and are fully connected.
• 58:08
  They will not necessarily believe every message that they see.
• 58:11
  So they already have built in better mechanisms of finding or ways of of eliminating the falsehoods from the truth.
• 58:23
  Now, of course, the platforms themselves can build fact checking algorithms, which they do.
• 58:28
  So there is a lot of work on the computer science side and that should continue.
• 58:32
  I think that is highly urgent that we do that, that should improve.
• 58:37
  So I don't have details because I didn't look at the algorithms and how they are designed and how they work.
• 58:44
  But that is, of course, very good. And then regulation is the other thing that the politicians need to regulate the information that is being spread,
• 58:54
  because the newspapers that are sending out information and the news channels, they have very, very clear and strict code of conduct.
• 59:06
  They they know what they are allowed to do. And they are not just broadcasting anything.
• 59:11
  So we know when a news agency is a reliable news agent, but on these new information platforms and social media, we don't have these these mechanisms.
• 59:22
  So they are totally not yet regulated up to the level that it should be in order to
• 59:27
  prevent people from being influenced by fake news and taking it as a truthful fact.
• 59:35
  So, yeah, so there are many things that we should do.
• 59:39
  And it starts with education. I think these games that are on the market.
• 59:43
  So there's a lot that people themselves can do and making them aware of this is is also very good.
• 59:51
  OK, great, thanks. We have a there was a there was another question by Mike and Mike, Mike Wooldridge, which is similar, I think,
• 59:58
  to what you've just been answering, sir, about this sort of work on automated fact checking and fake news, fake news detection and so on.
• 1:00:11
  I think he's mentioning in particular using deep learning to detect fake news.
• 1:00:17
  So he's asking if you think logic can play a practical role in this as well?
• 1:00:25
  Well, yes, I think that I mean,
• 1:00:27
  logic is the science for truth plays an essential role where I would say that the machine learning algorithms they don't necessarily have.
• 1:00:39
  So they might be able to deal with a large set of data.
• 1:00:45
  But the way that they deal with that data is not necessarily that they need to have an exact item that is labelled as the true one, right.
• 1:00:54
  So logic inherently is about valid inference.
• 1:00:58
  The notion of validity and truth is built into the whole practise of how logic works.
• 1:01:04
  That is not necessarily the case in the machine learning algorithms, of course, for fact checking.
• 1:01:10
  They will have something that allows them to to find the source of the information that goes to to the the one that is labelled correct.
• 1:01:20
  But it will then be automatically from the outside.
• 1:01:23
  Still needs to be labelled as this is the correct one. I think that logic can play a role and it should play a role because for but
• 1:01:33
  that is more an answer for not only for fact checking in the fake news case,
• 1:01:38
  but for all of A.I. where you do want to make you want to give an explanation of why a certain decision is made.
• 1:01:49
  Right. And so the machine learning algorithms by themselves are not necessarily very good at exactly that point.
• 1:01:56
  So explaining how you get from input to output.
• 1:02:00
  So that whole track of the inferences is something where logic is very good.
• 1:02:05
  And so combining these two methods would be ideal, where you have one tool that can handle large datasets much better than what the logic side can do.
• 1:02:17
  But for the explanation part, according to me, logic is very good.
• 1:02:22
  So so ideally, you would build a new system that combines the two.
• 1:02:27
  I know that people are working on this, but yeah, that's actually a big challenge.
• 1:02:32
  So combining the old and new AI together. Yes.
• 1:02:37
  OK, thanks very much then. Well, actually, let me ask another question.
• 1:02:44
  This is, do you think there could be an application for research and market simulation?
• 1:02:51
  Um, I think so. I mean, so the one of the things that we are doing now is, for instance, using model checking to verify.
• 1:03:06
  So you use logic to specify certain properties of behaviour of the network, of the agents,
• 1:03:13
  of the beliefs that they have now, then you can apply that to a whole bunch of situations.
• 1:03:18
  And then the question is, OK, this gives us insight in the behaviour.
• 1:03:23
  And so you could use it, for instance, when when agents anticipate that you are going to buy the same type of shoes that as your friends, et cetera.
• 1:03:35
  So so if we can anticipate that a cascade is going to happen, then we can predict what the next fashion is going to be.
• 1:03:44
  Of course, it is very difficult to do that.
• 1:03:46
  Many factors play a role and we, the logicians typically abstract away and reason at a very high level of abstraction.
• 1:03:53
  But could the principle be used for that? Now, if you actually want an actual application, then I think so.
• 1:04:03
  Going back to the to the A.I. story, what you could do there is that in cases where you don't have a lot of data available and
• 1:04:12
  you want to reason very exactly about the behaviour and the specifications of the agents,
• 1:04:19
  you could do run a model tracking algorithm for that behaviour to check if it is true in a network.
• 1:04:26
  Right. So it will be much more accurate than some of the other algorithms that are out there.
• 1:04:32
  So then you would use some of the tools from computer science with the logical specification of these properties,
• 1:04:38
  and you would do the role model checking on these networks that can lead directly to some application.
• 1:04:45
  And you could. Yeah, it's not necessarily for predicting what the market's going to be or what will happen to cryptocurrency or with the next version.
• 1:04:56
  But so this is more on spotting the influencers or the bots in the network, for instance.
• 1:05:06
  Right. Yes, thanks. If I could add a question of I mean, I find this discussion fascinating on one of the other information,
• 1:05:16
  Kaskade example that you mentioned, but the models with communication and the way that this kind of defeated event sort of dynamic,
• 1:05:26
  dynamic, epistemic well, analysis with where you could model higher order beliefs in some sense,
• 1:05:32
  I'm following a comprehensive analysis that you still wouldn't kind of refute this or detect in some sense this information.
• 1:05:42
  Cascais, on the other hand, you know, we standing outside can see that there's something wrong here.
• 1:05:49
  So, I mean,
• 1:05:50
  does this suggest a further kind of level of analysis we could put back into a logic that would do a better job of detecting information flow?
• 1:06:00
  Cascades, Cascade's? Yeah, indeed.
• 1:06:06
  Of course. So if we would model the modellers that are modelling the phenomena so that some kind of.
• 1:06:12
  Yes, some kind of reflection, you need an extra reflection, which is not just what the agents that are making the decisions,
• 1:06:22
  but that is just the the ones from the outside who are looking at this and strategically may be reasoning about it.
• 1:06:30
  And we see this happening. Right. So agencies are looking at the formation of these cascades,
• 1:06:36
  because if I can predict the market and maybe just start market cascade so that everybody will buy the same type of shoes tomorrow,
• 1:06:44
  namely the grant that I might be producing,
• 1:06:47
  then I can make sure that I can manipulate the first agents in the group and hope that the North will follow.
• 1:06:55
  And I can set the trends right. So so these mechanisms mean that you don't understand how this works and there are models for it.
• 1:07:05
  But these are at a higher level because at the level of the agents themselves, they're higher order reasoning.
• 1:07:13
  Power is not going to break the code, but the ones from the outside,
• 1:07:16
  they can say they can see exactly where to intervene, where they could manipulate the formation of the cascade.
• 1:07:23
  They know exactly which information to drop out, at what point in order for the cascade to derail or to stop.
• 1:07:31
  So, yeah, so that we didn't do that in our logical setting. That would be very interesting.
• 1:07:36
  Yes. Shows how subtle the and rich the sort of modelling capacity that one is led to and analysing these situations.
• 1:07:47
  I think also having to make a comment, taking out what's already been said.
• 1:07:51
  We now see there's a whole genre of books that because there's been a lot of statistics around in the public domain with the pandemic's,
• 1:07:58
  we see a whole chapter of books covid trying to interpret this for the public at large.
• 1:08:03
  It occurs to me that maybe something else also brings in this epistemological
• 1:08:09
  ideological aspect would be a very valuable resource for the for the general public.
• 1:08:14
  So maybe in some of the spare time for all the other things that you're doing, you should think of doing that.
• 1:08:24
  OK, so do you have any any other questions for Sonia?
• 1:08:34
  OK, if not that I don't see any right now, then yes.
• 1:08:40
  Well, thank you very much again for a fascinating lecture. And it's really been shown that logic is going into these new and actually very central
• 1:08:51
  and important areas and in current life and doing so with a lot of incisive ideas.
• 1:08:58
  So thank you very much, Tanya. Thank you for the invitation.
• 1:09:02
  Was almost a pleasure to give this lecture. OK, so that will conclude there.
• 1:09:08
  OK, thanks, everyone.