Virtual HLF 2020 – Talk: Andreas Matt
This is a modal window.
The media could not be loaded, either because the server or network failed or because the format is not supported.
Formal Metadata
| Title |
| |
| Subtitle |
| |
| Title of Series | ||
| Number of Parts | 19 | |
| Author | ||
| License | No Open Access License: German copyright law applies. This film may be used for your own use but it may not be distributed via the internet or passed on to external parties. | |
| Identifiers | 10.5446/49569 (DOI) | |
| Publisher | ||
| Release Date | ||
| Language |
Content Metadata
| Subject Area | ||
| Genre | ||
| Abstract |
|
00:00
MereologyImaginary numberPresentation of a groupComputer programmingMathematicsMeeting/Interview
01:11
Game theoryObservational studyDivision (mathematics)Ocean currentProjective planeOrder (biology)MathematicsPhysicalismLine (geometry)Game theoryWage labourProduct (business)Series (mathematics)Computer programmingMereologyWeightGoodness of fitCartesian coordinate systemNumerical digitGradient descentMeeting/Interview
04:46
Wave packetNumber theoryStatistical hypothesis testingMathematical optimizationMaß <Mathematik>Convex hullGradientCircleNumerical digitFunction (mathematics)Greatest elementTerm (mathematics)Different (Kate Ryan album)WeightOrder (biology)Multiplication signMereologyConfidence intervalRight angleGradient descentStandard errorGame theoryCircleArithmetic meanWave packetSet theoryHydraulic jumpMaxima and minimaMathematical optimizationNumerical analysisMathematicsGrothendieck topologyModulformStudent's t-testWage labourFunctional (mathematics)Exclusive orMaß <Mathematik>LeakMortality rateSparse matrixArrow of timePressureProgram flowchart
10:05
Imaginary numberMathematicsConvolutionCausalitySet theoryPresentation of a groupArithmetic meanOpen setMoment (mathematics)Real numberVapor barrierUniverse (mathematics)
12:38
MereologyUniverse (mathematics)Meeting/Interview
13:14
Duality (mathematics)Functional (mathematics)Convex setGrothendieck topologyMeeting/Interview
14:20
Insertion lossFunctional (mathematics)Maxima and minimaMeeting/Interview
14:37
Game theoryExpected valueGradientLocal ringPoint (geometry)Term (mathematics)Gradient descentMaxima and minimaHydraulic jumpMeeting/Interview
15:15
Game theoryMaxima and minima1 (number)Meeting/Interview
15:37
Category of beingModulformAnalytic continuationDifferent (Kate Ryan album)Group actionState of matterExtension (kinesiology)Game theoryContent (media)Ocean currentPhysical systemHydraulic jumpMereologyHeegaard splittingMeeting/Interview
17:55
Duality (mathematics)1 (number)Different (Kate Ryan album)WaveCartesian coordinate systemFunction (mathematics)Meeting/Interview
19:24
SpacetimeTime domainExpressionIdeal (ethics)Dimensional analysisMeeting/Interview
20:28
Group actionSpacetimeMeeting/Interview
21:10
SpacetimeReal numberGoodness of fitMeeting/Interview
21:31
Lattice (order)
Transcript: English(auto-generated)
00:27
As the last speaker of today, it's my pleasure to introduce Andreas Matt, co-founder of Imaginary, a company specializing in creating exhibitions and other outreach material for all issues related to mathematics.
00:42
Today, he will introduce the exhibition IMAI, a very nice palindrome by the way, which is part of the satellite program of the Virtual HLF. After the pre-recorded presentation, there will be a Q&A session live, and you can submit your questions as usual through the chat channel of the HLF app.
01:06
So, we can start the recording.
01:34
Andreas, and I'm very happy to introduce you to IMAI, Explaining Artificial Intelligence,
01:42
an exhibition project that was supposed to start this year in a physical open source traveling exhibition format, and at the end became a novel digital online format that I'm happy to introduce you today. And it will be the digital exhibition of the Virtual HLF.
02:03
And the idea, as the title says, Explaining Artificial Intelligence is to introduce a general audience, like a really general audience to some of the core methods, like the core basic technologies, the mathematics and computer science behind these AI methods,
02:22
and maybe going a little bit away of the whatever general hype of AI or the marketing part or the application part to really introduce some of the techniques. And the idea was to use some novel formats. So, you will find, I don't know, interactive videos where you can play with apps and listen to an explainer,
02:46
or you can read a comic, you can build your own AI just with paper cups and pencils. And we developed a series of, I don't know, call them apps or interactive experiments, where you can play a game to explore how gradient descent works
03:03
or interactively change the weights in the neural networks or play with a robot doing reinforcement learning. And some of the exhibits you can experience in this digital format, the others will open in the physical exhibition that will probably start next year,
03:21
under an open source license. And yeah, I'm very happy to introduce you to some of the exhibits. If you look at them, maybe, of course, take the eye, your researcher's eye, and look at these exhibits with, I don't know, with a curious researcher mind, and give us feedback, I don't know, connected to your own research and maybe add your own ideas.
03:47
And, but I would also look at it, I don't know, with the other eye, which I would call the outreach eye, the communication eye. I think it's so important to introduce the, I don't know, the core mathematics of AI in a way.
04:03
So look at it with a curious outreach eye to look at the formats and to see if this could work for general public. In the next couple of minutes, I'm going to introduce you to two of the exhibits. And we, like the team of imaginary, will be with you during the HLF in the so-called live sessions.
04:23
You can find it in the program in the breaks. So some of the authors of the exhibits, we will be there to answer your questions, to listen to your new ideas. And yeah, I wanted to thank the Carl Zeiss Foundation, who is the supporting body behind the exhibition.
04:41
And let's now jump into some of the exhibits. Hi, I'm now inside the iMAI website. The website is i-m.ai. And you can find here five, the blue buttons are interactive trails, and the yellow button is this virtual interactive tour where you can dive into,
05:05
yeah, hi, Antonia. So dive into some of the apps and Antonia is going to guide you through it. I'll jump to one of the first apps. It's called Neural Numbers. It's also an exhibit of the exhibition. And it's a standard example. I would call it, it's called the Hello World example of AI.
05:24
So it's the non-trivial task to recognize handwritten digits. So here I can just hand write a digit. And on the right side, you see the guess of the AI, of the machine. And in the center, you would see the confidence bars.
05:42
So the confidence bars, of course, they will change if I change the number. So now it's very confident that this is five. This is already a trained network. So it's quite okay. It's quite good. Of course, you can try to trick it and see, okay, now this becomes an eight,
06:01
somewhere between an eight and a nine. Interesting. But you can also start training the network. So here we have a network. And this is a non-trained one. So if I give it an input or just randomly say any number. And now I have a big set of training data. That's the good thing with this MNIST dataset.
06:21
So you have 70,000 images of labeled handwritten digits. You see them here on the left side. So labeled means I have the image of a two and I know that it's a two. And this training data I'm going to feed into the network. And the network is updating or improving itself in order to predict the output.
06:41
So it changes numbers inside. And if I stop it here now. So I can see now it's trained with 4,300 images. Live trained right now. And I can see now it would already recognize a three. You can see here the 90% accuracy. That's an interesting thing that you would test this network
07:00
on 20% of the training data that you don't use for the training to see if it works fine. We have here also an advanced mode. So that's for you to play around if you want to change the network architecture, activation functions, or even use a different optimizer. You can also see what would be the difference to train
07:21
with a different network architecture and to see how long it would take. It's faster or more accurate. Another exhibit is the so-called gradient descent game. And for that, I need my pirate hat. So it's a game where a pirate hid a lost treasure,
07:45
a lost mathematical treasure somewhere on the bottom of the sea. And it's really hidden on the very bottom, in the lowest part of the sea. And now, a couple of hundred years later, we have these expedition boats. And we can, on certain spots, I can choose a spot,
08:01
I can lower a probe. So this is like a probe. And I get, as a feedback, I get some information on the shape of the bottom on that part. So I lower some probes. And now I have certain trials. And the idea is, can I find the lowest part?
08:21
So it's very interesting. I'm not explaining a lot, and you can already imagine what would you do, where would you go in order to find the lowest part. Of course, you would look at the slopes, at the gradients here. Okay, I'll try here. Maybe this is already the good part. Yes, I found the treasure.
08:41
And I have some nice instructions for squaring the circle. And we can play several times. And it's very interesting if you look at it in terms of finding out what is gradient descent, what is the algorithm, where do I search, how do I change my step size, can I, I don't know, be stuck in a local minima.
09:04
And it's also interesting if you ask students, we did this in many workshops, can you draw a very difficult kind of floor? I don't know, maybe something like this. So the treasure would be hidden here on that side. Or you make a floor which has a lot of local minima.
09:20
You do something like this. And of course, it might be very difficult to find the treasure exactly here. And yeah, it's interesting because this is connected to neural networks in terms of connected to the error function. So what you do here is you minimize the error.
09:40
You try to find a spot of your weights, of your weights and biases that exactly lets you find the minimal error. So the lowest part of that error function. All right, so I invite you to play. You can also play in multiplayer mode. And yeah, I'm going to jump back to the big screen.
10:05
You have seen in these experiments that there are a few things in common. So first of all, they're kind of entertaining or in a way lightweight, but at the same time, they have some core mathematics
10:22
or computer science algorithms in there. And one thing is very important is, I would call it interactivity or creative interactivity. So you can really, I don't know, change parameters, play with these apps, get immediate feedback. I don't know, break things and really experiment on your own.
10:40
And I think that's very important to explore new technologies. Now we're going to have a 10 minutes Q&A session. So I'm open for your input. And we are trying to extend this, let's call it a set of open source tools to communicate AI.
11:00
So I'm really looking forward to new ideas, new exhibits on whatever causality or some convolutional networks and late research things. So thank you very much. And here you are now in the Q&A session.
11:58
I think we can start the Q&A now.
12:04
First of all, thanks for the very lively presentation and the overview of your exhibition. It's really, well, creating a lot of interest, I hope, and gaining you a lot of visitors. The first question I got from the audience
12:22
is a very simple one. Is the exhibition available to the public? Can I direct people, friends to the exhibition somehow? Yeah, so the first physical exhibition is planned to open in May 2021.
12:41
And it will open in Jena, Germany. And with that moment, it will also open under an open source license. So which means that if you cannot come to Jena, it will be shown at the university in Jena. You can also copy the whole physical exhibition and show it in your own city or country or university.
13:03
And until then, I invite you to visit the digital version, which is part of the exhibition. It's not a full exhibition. And when you say you can copy the physical exhibition, what does that mean really? I mean, how much of an effort would that be?
13:22
Do you have an idea? Yeah, I mean, there are about 15 exhibits. And depending on the technology you have available, some of the exhibits are, they run on more or less good computers and you need touch screens. So you can just show them running on a touch screen. Some others would need, I don't know, camera
13:42
or some physical hands-on items. But in general, if you have some technology on site or you have some budget to rent or purchase some technology, the exhibition is made in a way that it's not too complicated to copy it. Of course, you can always build a super,
14:02
I would say a super fancy case for the exhibit. But if you want to show it in a school, you can just use even normal computers. If you have touch screens, it's even better, but it would also run using mouse or keyboard input. Okay. The next question, I read it.
14:23
And if you want, you can take on, but otherwise I would suggest we pass it on to David Silver. How can we ensure that minimum loss value is achieved if the loss function is non-convex? Yeah, so as I understand non-convexity,
14:40
you have a lot of local minima. So it can look like this. And if you get stuck in a local minima, I mean, you can look at play the gradient descent game. You would at one point have to make a big jump. So in computer science, very often it's called exploration versus exploitation.
15:02
So exploitation would be to go down, like follow the gradient going down. But if you're stuck in a local minima, at one point you have to apply a certain random big step and you do exploration in terms of jumping to another, hopefully the global minimum, or you will find another local minima.
15:21
Okay, thank you. Next one, thank you for the teaser. Do you provide explanations in your exhibits and which ones work best for the audience? I'm not sure which audience is referred to, but let's assume it's the HLF audience. So it's very interesting, the question about explanation.
15:40
So it's a very, there is, first of all, there's a big difference if you have an exhibition versus a museum. So in an exhibition, very often you have a human explainer on site, and that's still the best. The best is to have somebody who is motivated, who knows how to deal with human beings, who knows how to explain, who knows the topic,
16:01
and who is there to guide you, to ask questions, to answer questions, to be with you, to explore the content. Of course, very often you write text. So text for many people is still a very good and fast way to gain knowledge. Usually, depending on the age group and target group, texts nowadays in exhibitions is not,
16:22
people don't read a lot of texts. So they would read a few questions, a few things, a few paragraphs, but not a lot. So what we do in some exhibitions, we provide a booklet that you can even take home. So you have a small booklet that you can take from exhibit to exhibit, and you have your individual kind of guide
16:41
that you can take home and continue reading. So that's a good system. But I would say the most important thing is the human explainer there. For the digital format, we are just now trying to develop, I would say, new ways of combining the games with the text.
17:02
So if you look at the website, we call them trails. We have these interactive parts, and then you have some text, you have a video, you have some text, and that I think also works quite well. So it's a common format on the web where you have longer texts, but it's kind of split up with apps and videos in between. And we explore on the website,
17:21
on the IMAI website, a very new format where you have the app running together with an explainer. I showed it in the video, Antonia, our team member. So it's a prerecorded video, but the video that controls the app. So if you can still use the app and jump forward and backward, and depending on where you are, Antonia will tell you other things. So it's a kind of interactive video connected to the app.
17:43
And that's something that I like a lot. I think there's a lot to still explore how to add the explanation part to content. Next question. What are the natural language applications of IMAI?
18:01
We have two exhibits connected to natural NLP, which both are not yet online because they require some special hardware. And we have one exhibit, it's called Talk to Me. And yeah, you have a microphone and you just talk to the exhibit, you say something, and then you would in real time see
18:21
how a convolutional network would analyze the audio. So, and transform the audio like the wave or version of the wave into text. There's two languages that you can choose.
18:42
You can also try to speak German for the English neural network. And then you would see, you can see the text and you can see all the different steps. Like it's also visualized in a way from the wave file to the final output cleaned up text. And it's really interesting to see all the steps in between.
19:02
So it's also an interactive app, but it only runs on a good hardware. So we could not bring that exhibit to the web yet. Next question. Really cool. That's not the question, that's the statement. Have you thought about or implemented a visualization
19:20
for the Bayesian posterior? No, not yet. And it would be really nice. I mean, the concept is the idea of IMEI is to look at the big column, the big main tools of AI. So this would definitely enter that domain and it would be really nice to have something.
19:42
And I think it also combines to the, maybe to the next question. But like, if you have an idea, we will be doing the whole HLF. We will be in the afternoons. There are two sessions where you can find us, like not only me, but also team members of IMEI, to discuss ideas, to look at proposals,
20:04
to look at your own research, and to see how we can transform it in an interactive app or in an exhibit that we can then implement. The physical exhibition, maybe that's also interesting to mention, will have a space for, we call it a research space. So they're the pre-made exhibits,
20:20
but there's a space where we can still explore new things. So we are very open for new input on that side. That essentially already answers the last question, but I read it anyway. Thanks for your nice idea. How can we send you our ideas about creating new games?
20:40
Yeah, as I said, I mean, you can send it to us via email, but if you're here at the HLF, let's meet. I think that's also the purpose of the conference, so it would be really nice to meet you. Tomorrow at 3, we have the first so-called live session on IMEI. So if you could join us tomorrow,
21:01
the person who asked that question, that would be really nice, and we discuss all types of ideas. Maybe bring them along and let's look at them. Okay, great. Thank you very much, Andreas. It was a very nice presentation, and as you can see from the questions, people are really excited about your exhibition.
21:21
And I wish it a great success both in cyberspace and in real space. And I wish you a good evening or whatever is left of it. And to the audience, we will close down today's sessions. Thank you for joining us. You may want to hang around in the virtual space a little bit
21:41
and meet other participants and just socialize, exchange ideas, whatever. But make sure to be back tomorrow at 4 p.m. CEST for the second day of the virtual HLF. Thank you and good night.