Introduction

So welcome to the final presentation of this evening. My name is Charlotte Gerritsen. And as you can see, I'm, as my sponsored shirt says, from the Vrije Universiteit in Amsterdam, where I am an associate professor in AI, and I'm the head of the AI and behavior group.

Research Focus

So within our group, we focus on using AI methodology to study, to understand, to predict, prevent and support behavior and behavior change. And we do this on three different domains.

So health, both mental and physical health, sustainability, and criminology. And that is my expertise.

So I have a background in law and criminology, different PhD in AI, and all the research I do is focused on that intersection.

Domains of Application

So like I mentioned, we have four different fields within the group.

And for understanding criminal behavior, we use AI to model and simulate behavior. So this could be knowledge-driven or data-driven. Knowledge-driven would be based on theories from sociology, psychology, or criminology. And data-driven would typically be based on large data sets and would mean that you use machine learning.

And when I was discussing this presentation with my colleague, he said to mention that explicitly mentioned that we do not use chat GPT LMS or gen AI, but more traditional AI methods.

So 1to understand criminal behavior, an example of one of my colleagues is that he built a VR environment where he had people burgle houses in the neighborhood. And he had all kinds of people doing that, also people who are imprisoned. So he could really see how professional burglars were different in their behavior than random people who tend to move differently in a house of someone.

So that would be an example of understanding.

Understanding Criminal Behavior

Then predicting is something I will come back to later, but about who, when, and where crime will be committed.

We build a lot of VR-based training environments where people would be able to interact and train their own behavior to be ready as far as you can be when you encounter a criminal situation.

Practical Implementations

So I will mention three examples. One is for elderly. We had a VR training environment where elderly could train how to deal with scam artists on the doorstep. So if they are at home, someone comes to their home, they open the door and someone tells them, oh, can I, I need your PIN code. I need to... take something from your house, can I go to the restroom, whatever. And they would practice during these scenarios what to do and how to be assertive. So we incorporated an assertive measurement where they could actually speak to the system and they would get feedback on how assertive they were.

1Another example is aggression de-escalation training for people in public transport. People encounter a lot of aggressive behavior. How to detect the right type and the right approach was something that we did. We did that together with the GVB and the RET. And actually, at the RET, they implemented that in their training environment. So everyone works with this.

And the third... example I wanted to give was for people working behind the cash register and for instance, Albert Heng. They typically don't get the training on how to deal with a robbery, especially because the turnover of the people at the cash register is quite fast. And usually only management level gets the training. It's quite expensive. And we built an training environment where they could train at home how to deal with these situations. Because people tend to want to protect the company and keep the money, but the actual answer is to give it away as soon as possible. Most companies are insured and then you will be safe at least.

Support and Mediation Environments

And finally, support. So that is a project that we're currently working on where we are developing a support environment for people who became victim or even offenders who want to participate in mediation.

So after an offense already took place and where one of the parties is not interested in joining the mediation. So you need two people to do that.

And we built a virtual environment where people are able to interact with someone. virtual character, but to express their emotions and to at least say what they want to say to that person or someone that resembles that person.

Predictive Policing

So while I can talk about this for hours and weeks and a lot, I only have 15 minutes, so I will focus on the prediction here. And to be more specific here in predictive policing, so what is predictive policing?

That's the use of new analytic tools and large data sets by the police in order to find patterns and forecast future offending. So who will be most likely an offender or a victim?

And for, well, at least people of my age, they might have heard of the Minority Report movie.

I always refer to this during lectures and then you can see if you're still, well, around the same age as your students. Well, not anymore. So Minority Report and Persons of Interest. Minority Report is a movie, Persons of Interest is a series, and they are both about predictive policing.

In both cases they use, so they are science fiction movies and series. And in both, they try to predict who will commit a crime and prevent that from happening.

And like I said, it was science fiction in 2002, it was still science fiction in 2011, but it's not science fiction anymore. So currently there are quite a lot of predictive policing initiatives over the world.

Prediction of Crimes and Offenders

You can make the distinction between the prediction of future crimes So can we forecast where or when crimes are more likely to occur?

And predicting of future offenders and or victims. So who will most likely be at risk of being one of those?

And using large data sets to make these predictions, it has pros as well. For instance, prevention, so making sure that crime does not happen. But also, it's very effective if you know that there's most likely be an offence in this street today, you can just direct the police to this area instead of an area in Amsterdam where no crime will occur.

However, there are a lot of cons as well. For machine learning, you need large datasets. And these datasets are most often biased, especially within the police domain.

A lot of people, well, police tend to focus on specific characteristics of people. They, for instance, focus on young males, and then there are no females in the data set, or people that live within the city, a certain area, so then they lose. Very important information to be able to make an objective prediction. This leads to discrimination.

Challenges and Concerns

Of course, there are privacy issues. Do you want your information in the system? Accuracy, how about false positives?

If the system thinks that you will commit a crime in the nearby future, should you be arrested for that? And what if the system was wrong? Could be.

And a final thing, it is based on data from the past. So if you were committing crimes during your teenage years, does that mean that when you're 40, you're still most likely to commit a crime or not? Yes.

Crowd Management and Event-Based Predictive Policing

And then now I want to focus on a project that I did for the last five years, which I know for those of you in industry sounds like very, very long within academia that we're just getting started.

And it's about a specific type of predictive policing. It's event-based predictive policing, also known as crowd management that we're focusing on.

So, just by a show of hands, who of you have ever been to Amsterdam during King's Day? to one kind of sporting event, could be Barcelona soccer, could be anything, soccer, football, yeah. Or to a festival or concert. Yes, also most of you? Good.

So these events where large crowds gather together are supposed to be fun. People go there to have a good time.

But there can be just only one small thing that leads to, well, serious consequences, riots, people getting hurt, people trying to escape.

Yes, so that was the starting point of a project that we've been working on for the final five years. And I have a short video to show. It will work.

Research Lines

So as you can see in the movie, we have two different research lines within this project. The first one is emotion contagion.

So which emotions are actually present in the crowd and how can we determine where a different emotion is present. So if you are at an event that is supposed to be fun, we want to see where people are not happy or not happy, aggressive.

Can we determine subgroups within this group? And where are they specifically located within the area?

And we do this by automated analysis of text, video, and sound. And now you might think, okay, but how about privacy or checking everything about us at the event?

But we do this on a higher level. So we don't actually, for instance, use words. and conversations, but we use the tone and the volume to be able to do that. We are able to distinguish different sounds within a crowd based on that.

And then the work of Erik, the second one, is about emotion quantation. So how does the emotion spread from one person to another?

And we do this by using modelling and simulation. And to be able to manage this, so to send the amount of guardians that you have located at a certain event to the right position and how to deal with that, we developed a virtual reality-based training where you could actually detect emotions and learn how to deal with that.

Conclusion

So to conclude, Predictive policing, especially within the crowd management scenarios, has a lot of benefits, but there are also downsides.

When you look at the abstract level instead of the individual level, at least you take care of some of the main problems that you can have with these systems. A thing I wanted to mention is that it might not be necessary to actually police, but to monitor.

So if at a large crowd event people know that something will happen, that doesn't mean that all the guardians necessarily need to go to that location and stand in that corner to look at the people over there. But they can just monitor via the cameras they have and the monitors they have in the control room.

then something that is very important and that we, of course, when we started, AI is very fast in developing, so things change over time. We started with the best intentions, but after working on this for years, we of course also realized that you can use the things that we develop for not intended purposes, which is very important to realize.

We know that as well. We can detect unhappy people within a festival, but it also means that in China, there's a system that actually detects students that are not interested in a lecture. what they do with this information. I don't know, but yeah.

Well, I think we're trying to see a deviant behavior of people behaving differently than the rest. Um, And then finally, that follows from the previous point, is that it's really important to have strict regulations.

So I assume many of you are aware of the AI Act. The AI Act, they identify different tiers of risks where depending on how you use your crowd management. You are in tier three or four, so the more risky areas, which is very important to realize.

And of course, that leads to important restrictions to your work or the applications as well. These are my contact details. If anyone wants to discuss anything, even after today, just send me an email.