Police departments around the world are using this startup’s AI to predict future crime.

Mami Kajita, founder of Singular Perturbations, explains the success of their models, the public reaction to the technology, and how the physics models of glass transition lead to a crime prediction AI.

We debate the future impact of crime prediction technology, and we also talk about how researchers and entrepreneurs can better connect and collaborate.

It’s a great conversation, and I think you’ll enjoy it.

Podcast: Play in new window | Download | Embed

Subscribe: Google Podcasts | RSS | More

Show Notes

  • Telling police what future crime is likely to occur
  • Who else, besides the police, can use these tools
  • How the physics of glass transition lead to crime prediction
  • How to sell software to the police (and other government agencies)
  • Real world trials led to a 68% decrease in crime
  • What data go into Crime Nabi’s models
  • The public reaction to future crime prediction
  • Unintended consequences and and the future of crime prediction
  • How founders can find mentors and advisors
  • How researchers and entrepreneurs can better connect and collaborate

Links from the Founder

Transcript

Welcome to Disrupting Japan, straight talk from Japan’s most successful entrepreneurs.

I’m Tim Romero and thanks for joining me.

Today we’re going to talk about predicting future crime, and not in terms of 1950s science fiction, but in terms of real software being used right now by police departments all over the world. 

We talk with Mami Kajita of Singular Perturbations about their Crime Nabi AI, and how this technology is starting to change policing. In real world use Crime Nabi has already resulted in crime reductions of over 50% in areas where it’s been tested around the world. 

In our conversation, Mami and I dig into these numbers and we talk about the somewhat surprising inputs that go into training the Crime Nabi AI. 

And of course, we also talk about the very real potential dangers for misuse and what Singular Perturbations is doing to make sure this technology is a force for good. 

Along the way, we talk about how founders can find good mentors and advisors, the proper balance between research and sales, and some really good advice about how to sell to national governments as a startup. 

But, you know, Mami tells that story much better than I can, so let’s get right to the interview.

Interview 

Tim: So, cheers.

Mami: Okay, cheers.

Tim: So, I’m sitting here with Mami Kajita, the founder and CEO of Singular Perturbations, the AI for Crime Prediction. So, thanks for sitting down with me.

Mami: Yeah, thank you so much. I’m very honored to be here, and thank you so much for this opportunity.

Tim: I’m glad to have you here. In the intro, I explained a little bit about what Singular Perturbations does. But I think you can explain it much better than me. So, what does Singular Perturbations do?

Mami: We predict future crimes in using AI technology and we provide operation management services for police departments and local governments. And the name of our product is Crime Nabi.

Tim: So, you are telling police departments where future crime is likely to occur?

Mami: Yes. Using this technology we can provide the area where the risk is higher than the other area.

Tim: Okay. And how do they use this information? What do they do with it?

Mami: We provide operational management services in the police department, and there is a team who patrols outside and in Japan, and many police departments doesn’t use crime prediction technology before patrols. They have not so much established plan.

Tim: So, the police departments are using this predictive technology to decide where to send patrols?

Mami: Yes, yes.

Tim: Okay. What kind of predictions does it make? Does it predict the type of crime or just the level or…?

Mami: Our crime prediction technology predict future crimes, when and where, which crime type likely to occur.

Tim: Looking at your website though, you also have an app that individual citizens can download?

Mami: Yeah. This mobile application have users of citizens and in Japan, in local governments, they have a team who patrols outside using blue patrol car, and we provide our crime prediction mobile application for them.

Tim: Oh, okay. So, it’s not individual citizens, it’s still the government? 

Mami: Yes. 

Tim: It’s just not police?

Mami: It’s not police department, but many of blue patrol cars are operated by local governments or citizen.

Tim: So, it’s similar to in America we have neighborhood watch. 

Mami: Yes, neighborhood watch.

Tim: It’s a similar program to that. 

Mami: Yes. Yes. 

Tim: Let’s dive deep into the technology, the business model a little bit later. And let’s talk a little bit about you. So, you founded Singular Perturbations back in 2017. And back then you were a researcher at the University of Tokyo. So, did it start out as a research project? How did you get the idea to start this? 

Mami: I have a background of theoretical physics. My research topic is about glass transition. That topic is very different from the crime prediction technology, but I worked for university for several years, and after that I have to move to Italy because — so my husband have to move to Italy, so I…

Tim: Went with him. 

Mami: Yes. At the time I quit my research topic and I was thinking about my future. And my research background is theoretical physics, but I think that theoretical physics is a very basic topic. And theoretical physics people have a potential to make in models.

Tim: I’ve always thought, I majored in physics myself, but I’ve always thought that physics is kind of the liberal arts of the sciences.

Mami: Yes, yes.

Tim: You understand a little bit of everything but until you get a PhD, you’re not really qualified to do anything in particular. So, the mathematics and the modeling for the glass transitions, is it similar?

Mami: Yes. And very similar. Using this background, this kind of basic liberal arts background helps to create a new business. I think that there is a many similarity between researcher and entrepreneur.

Tim: I think so too. And I want to talk about that, but I’m really curious, what is the connection between the physics of glass transition and crime prediction? It’s not an obvious link.

Mami: Yes. Yeah, yeah. Eight years ago I started to live in Italy in Bologna. And I have experiences to be a pick pocketed several times. And at the time I wanted to start some new services using theoretical physics background. So, I studied to develop mobile application and I collected crime data. And then I developed a mobile application, which shows when and where the past crime occurred. And also I found a theoretical paper, which is about crime prediction technology. And this paper was written by theoretical physicist. So, I started to analyze the crime data using this kind of technology. And then I invented a new algorithm to predict future crime. And this achieved very good accuracy.

Tim: Were you a researcher while you were living in Italy, or was this just a side project you were doing for fun?

Mami: Ah, this is for fun.

Tim: No, that’s awesome. I love that.

Mami: Yeah. And after that, I sold this mobile application to a Japanese company and then I wanted to start a new crime predictions project.

Tim: So, after you sold that mobile application company, what brought you back to Tokyo? How did you start this startup?

Mami: After I came back to Japan, I entered a company where I sold my application and I tried to make a business using crime prediction, but it has several problems. So, I founded my company by myself after that. At the time, I was only visiting researcher at University of Tokyo, and I met one policeman. He’s very interested in my crime prediction technology, so he can invite me to a project of police departments.

Tim: This is interesting to me. So, when you went to the University of Tokyo as a researcher, were you going in thinking that, okay, I really want to start another company now? Or were you thinking, okay, I really want to get back to theoretical physics research?

Mami: After I have experienced long time research, I was very devoted in that topic but I also think that maybe I can help some people, but the impact is not so big. But using my background, I can make more big impact using this kind of technology. That’s why I was interested in the development of mobile application and the foundation of my company.

Tim: I think that is one of the most exciting things and the most addictive things about startups and many people discover it when they first start how valuable what you make is to your customers and how appreciative they are. And that’s something you usually don’t get in the business world or in as a researcher. 

Mami: Yeah. Yeah.

Tim: That’s nice. So, you’re still doing fundamental research and that was a NEDO grant, right?

Mami: I guess NEDO grant, yes.

Tim: So, one of the things a lot of researchers struggle with is the pull between doing long term fundamental research and having to sell this much product this quarter to make sure you’re on track to get your next funding round. How do you handle that conflict?

Mami: Yes. At the first, when I found my company, we can only provide crime prediction report, but this type of report doesn’t change the police department’s operation. So, we want to change the operation.

Tim: Well, let’s talk about some of your successful trials and proof of concepts you’ve run so far. You launched a project in Adachi ku in 2020 in Nagoya last year? So, what were the results?

Mami: Yeah. In Nagoya City introduced our product and many other local governments and the police departments started proof of concept. And now we started proof of concept in Latin American countries. And after this proof of concept, the number of crime decreased.

Tim: So, if I understand correctly, the way the proof of concept runs is you’ll work with the police department, analyze existing data, recommend new patrol routes based on this data. And what do they measure? What determines if it’s a success or not?

Mami: Yeah, yeah. There is a ways to measure impact. One of them is to measure number of crimes during the POCs. And the largest way is AB test. We prepare two areas where the number of crimes are very similar, and we provide our Crime Nabi application to one of them for the first month. And after one month, we provide Crime Nabi application to others.

Tim: Okay. And the idea is that because the police and patrols are being used more effectively, that’s what leads to the decrease in crime?

Mami: Yes. Yes. And after the POC, we can analyze the area where the team patrols and other area where that team didn’t use our mobile application, and then we can compare the effect.

Tim: That makes a lot of sense. And so what was the result? What sort of a decrease in crime did you see?

Mami: We found a decrease. This experiment ended last September. So, very new results.

Tim: So, what sort of a percentage decrease did you see?

Mami: It depends on the area, but the number of crimes decrease by 68%.

Tim: 68% drop.

Mami: Yes. But there is many reasons, so we cannot divide.

Tim: Oh, no, no. But that’s amazing. 68% reduction is astounding. Well, what types of crimes are we talking about here? Is it like graffiti type crimes? Is it like pick pocket crime or…?

Mami: Pick pocket type and this type of theft.

Tim: Okay. That’s a bigger number than I expected. Let’s dig into that a bit. Is that a result of increased patrols? It’s the same number of patrols going on in these areas. 

Mami: And this result is very analysis. So, we have to analyze more.

Tim: No, I understand. I mean, I’m not going to argue math with a theoretical physicist.

Mami: So, there is many effect from that patrol. Maybe it’s some number of crimes move to the area outside.

Tim: What goes into the AI model? What is the data you’re looking at?

Mami: Crime data changes depending on the police departments, especially in Honduras, the police departments are very, very interested in the homicide. So, we are focusing on homicide data in Honduras and in Brazil many police departments are interested in safe type crimes.

Tim: But in terms of like the training data, what type of data do you put into your model?

Mami: We use many data, sometimes past crime data and also land-use data. And sometimes we use satellite use data.

Tim: So, land-use data means like whether it’s residential or warehouse or housing.

Mami: Yeah. How the area is used, for example, with houses or the shops in the kind of shops. All kinds of image data we can put into our model.

Tim: In another interview, you also talked about some other data you were using weather data.

Mami: Yes. Weather data. Yes.

Tim: Traffic?

Mami: Yes. Traffic data, yes, we can use, but now our system doesn’t include them. 

Tim: I read another interview. You mentioned you were actually using Twitter posts.

Mami: Ah, we tried some years ago, yes. And it helps in some cities. 

Tim: Really? 

Mami: Yeah, yeah, yeah.

Tim: Why?

Mami: There is a paper that shows the Twitter data is effective for crime prediction because we use Twitter data with latitude, longitude, and sometimes with the data includes some traffic accident or some events.

Tim: No, I’m sorry, I’m just trying to wrap my head around why Twitter data has any predictive power at all. So, is it like people are tweeting about the accidents or the crime first? Or is it just where there’s a lot of people tweeting crime is low. What’s the influence there?

Mami: Ah, yeah, yeah, yeah. Many kinds of tweeted data is effective. For example, delay of airplane.

Tim: I’m sorry, why? I don’t see the connection.

Mami: Ah, sometimes this kind of accident changes density of people.

Tim: Oh, okay. And the density of people obviously has an effect on crime, particularly like the theft crime pick pockets. 

Mami: But the effect is not so big, but…

Tim: But it exists.

Mami: Yes. Yes.

Tim: Is that robust? Is that true in all countries?

Mami: Ah, we only test it in the United States, so.

Tim: Okay. Interesting. What’s been the public’s reaction to this? I mean, I’m sure you get compared to minority report all the time.

Mami: Yeah, all the time. 

Tim: What’s been the public reaction to it?

Mami: We provide services to governments now but we want to provide also for the general customers. We want to make a safe road using crime prediction technology. And we will start a new proof of concept in Uruguay from next spring. And also we want to provide our services for travelers.

Tim: So, with AI in general, people are both excited about it and kind of scared of it at the same time. Have you had any reaction from the public about whether this is good or exciting or scary or…?

Mami: Yeah. Some of people are scared about this kind of technology, and when we provide our crime prediction patrol route, some policemen have to move to the very dangerous area. And some of them are feared about this technology. We get this kind of interviews but I believe that is kind of technology helps to decrease crimes.

Tim: But isn’t there a chance of having this kind of negative feedback loop? So for example, you get more patrols in an area, so you get more arrests in the area. Which leads to higher crime in the area, I mean, how do you prevent that from happening?

Mami: Yeah. That can happen. And now our models doesn’t include this development, but this kind of bias is possible, so we have to put some randomness for the patrol route.

Tim: Oh, that makes sense. So, you do introduce some randomness, and then you look at the effect of that randomness to double check the accuracy.

Mami: Yes. Yes. And our crime prediction system, Crime Nabi also makes patrol route with the input of the point data. And police departments have to go to this kind of point and they can input that point to our Crime Nabi system, and we can make a patrol route. So, we don’t determine all of that pattern operation. We only support and policemen decide.

Tim: Well, I understand that, but as fascinated as I am with the technology of AI, we humans are deeply flawed. I think we want to believe computers too much. And every time AI’s been rolled out so for example, in America there’s, I don’t know if it’s AI, but computer systems in general that are being used to like set bail or for advising on sentencing. And human beings love to say, oh, well the computer did it, it’s not me, it’s the computer’s decision. And so you can say, well, the police have the final choice. But I think the natural human reaction is to trust the computer.

Mami: We don’t want to replace the human. We want to support because the accuracy is not 100%. This kind of technology can help people’s analysis and management people’s planning. But we should not decide total operation. We can make a heat map of crime prediction, and policeman can input the point where he have to go from this operation. He can put his intuition to the patrols. So, we want to blend the algorithm and the human intuition.

Tim: So, this been amazing developments just in five years. What does the future of this technology look like? If you look five years or 10 years ahead?

Mami: Yeah. We provide Crime Nabi to a local governments and the police departments now, but we want to provide crime prediction technology to general users like travelers or experts and crime prediction technology needs past crime data, but past crime data is not totally open. The case is very limited, the area is very limited. So, we developed a transfer learning algorithm, very new algorithms. And this enables us to predict future crime in the area where we don’t have crime data.

Tim:Okay. So, right now you’re focusing on government sales, police departments, security services. And you think the next step, a few years from now we’ll be moving more towards travelers and corporate travel programs, that kind of thing?

Mami: Yeah, yeah. There is a problem for travelers. They don’t have enough knowledge about security when they travel. And also they cannot call emergency call during travel. So, we want to connect governments and general consumers.

Tim: I think it’s amazing that your company was internationally minded from even before you founded it because of your experience in Italy. And do you find that your experience with foreign government sales makes it easier to sell to Japanese governments as well?

Mami: Japanese market of police departments are very special and very closed. So, maybe finding a partner company is more important to enter to Japanese government’s market. 

Tim: Yeah. I can see that.

Mami: This is very special. But in Latin American countries, that war is not so much higher than that of Japan. So, we provide our services to Latin Americans country. It’s not so much difficult, but in the Japanese market it’s very tough.

Tim: But let’s talk a bit about startups in Japan in general. Your jump from research to starting a company, I mean, that’s a big jump. What was the hardest thing you had to learn?

Mami: Yeah. There was two big difficulties. The first one is to know how to make business using technology, because I have only background of research, so I don’t have enough knowledge. So, at first I developed a mobile application by myself and I tested the users’ comments. After that I have to learn how to make a business model for this many advisors and mentors helped me out.

Tim: Okay. The mentors and the collaboration is so incredibly important.

Mami: Yes. Very important.

Tim: So, how did you go about finding these mentors and these advisors?

Mami: I tried many business contests and I got some prizes. And after that I met many mentors.

Tim: All right. I mean, that’s a great use of pitch contest.

Mami: Yeah, yeah. Yeah.

Tim: So, both Japanese government now and Japanese universities are talking about how important it is to take some of the research and the deep tech and turn it into startup businesses. How do you think we can encourage more researchers to start startups?

Mami: There is two type of researchers I think, the first type interested in more about research more than business. I recommend them to collaborate with some professional startup people. And the second type are interested in starting new business. I recommend them to found by themselves with our collaboration with founders

Tim: And pointing out that they can make a bigger impact.

Mami: Yes, yes. But many difficulty to find the co-founders for researchers. In Japan this kind of culture is not too matured yet. 

Tim: It’s getting better I think.

Mami: Yes. Getting better, yes. 

Tim: But yeah, the kind of open collaboration is still kind of new in Japan.

Mami: Yeah. Yeah. Yeah.

Tim: Even the kind of collaboration you were talking about before with working with criminologists in different departments, that’s also kind of new. Well listen, Mami, before we wrap up, I want to ask you what I call my magic wand question. And that is, if I gave you a magic wand and I told you could change one thing about Japan, anything at all, the education system, the way people think about risk, the way people think about their jobs, anything at all. To make it better for startups in Japan, what would you change?

Mami: For myself, I founded my company with my husband but the full committee member only me for a very long time to find members for my company. I have to struggle a lot.

Tim: So, why do you think it was so hard to find people? Were people too scared to take a risk or was it too complicated to understand or people aren’t comfortable with startups yet? What was the challenge?

Mami: I have research background and I have to learn many, many things to start a business. So, I wanted to meet more people who are open to such people. Cultural background of researcher and entrepreneurs are very different. But the mind is very similar. So, if they can collaborate, many big change in big impact we can make, but increasing, but not so much opportunity to meet and collaborate with researchers and entrepreneurs.

Tim: Do you think it’s just a problem of not enough opportunity? Or do people have the wrong idea about researchers?

Mami: Yeah, yeah. There is many gap for the mind, very similar, but they are big gap. For example, in many cases the way researchers think is very abstract, but the way entrepreneurs think are very concrete. So, communication is sometimes very different. But the mind is very similar and they also take risks and challenges a lot. So, to communicate we have to understand each other.

Tim: So, you would create more chances for researchers and entrepreneurs and people with different backgrounds to meet each other and exchange ideas? 

Mami: Yes, yes. Yeah. If I have this kind of opportunity, I can learn more quick and I can learn a lot.

Tim: Do you think that’s changing it? Because when I look at this, it just seems like it is improving both in universities. So, it’s going in the right direction. 

Mami: Yeah, yeah, yeah. 

Tim: Okay. Well listen, Mami, thank you so much for sitting down with me.

Mami: Thank you so much.

Outro

And we’re back. 

I hope you appreciate the effort involved in keeping my Minority Report references to an absolute minimum. It would be easy, it would be kind of fun. But what Singular Perturbations is doing here is something different and something very real. 

I mean, a 68% reduction in crime is too big to ignore. 

And yes, that number may come down a bit as Mami and the team analyze the numbers and isolate the causation from the raw correlations. 

But right now, this technology is being evaluated by police forces around the world and it’s going to be a game changer. Something this powerful is going to be used and used a lot. And on average the world will probably be a much safer place because of it. 

But these tools also have the potential to make life harder on some of us, particularly those who are already disadvantaged. So, far we humans don’t really have a healthy relationship with this kind of AI. We tend to praise the AI when it confirms our existing biases and overrule it when it contradicts them.

So, what’s the answer? 

Well, giving strong authority to low-quality AI will simply reinforce and amplify whatever biases existed in the training data. Disadvantaged groups will become further disadvantaged. However, giving strong authority to high-quality AI could actually help correct those biases. 

As Mami pointed out, adding randomization to patrol routes and rigorous analysis of crime data could show that setting aside our biases can lead to more effective policing and safer communities. Or alternatively, it could show that those biases are completely justified and that we should focus on incremental improvements. 

In the end though, it’s not a technology problem, it’s a human problem. Whether AI leads to greater equality and fairness or serves to further amplify our existing biases depends mostly on us being able to question our own beliefs and admit when we’re wrong. 

Now, the past 10,000 years of human history shows that we haven’t been very good at that, but maybe AI can help us get better. 

Maybe AI can help us become better humans. 

 

 

If you want to talk more about crime prediction or AI in general, Mami and I would love to hear from you. So, come by disruptingjapan.com/show199 and let’s talk about it. And if you enjoy the show, share a link online or just tell people about it. In this age of over the top hype, you’d be amazed how much power your honest recommendation has. 

But most of all, thanks for listening and thank you for letting people interested in Japanese startups know about the show. 

I’m Tim Romero and thanks for listening to Disrupting Japan.