Now, we're about to start the HLF poster flash, and we will have 30, or rather 29, as one unfortunately didn't make it to Heidelberg, 29 young researchers on the stage, and they have two slides and two minutes

to talk about their work, to share their project with you. So we kind of took the idea of an elevator pitch from investment to science, and give our young researchers the possibility to share their knowledge, as this is so important nowadays. I mean, you remember our hot topic on climate change.

We saw how important it is that science is in a position that it can help people to understand what's going on and what can be done, and therefore this is a wonderful way to practice, to break down your research into a very simple and short form.

Today is the European Day of Languages, and Europe is celebrating the diversity of languages it has. There are about 230 languages in Europe, and I thought it might be a nice idea to commemorate this European Day of Languages

by presenting the diversity of languages we have within our young researchers. So everyone who does come on stage will say hello in his mother tongue, and I guess that's a nice fit, because with this Bavarian evening, we have a lot of people in their traditional attire, and so you do not only look at these beautiful costumes,

you also hear the voices from around the world. So let's start the poster flash. There's one more thing you should know. We've clustered the coming talks, and our first cluster will be on machine learning, a field where we've seen promising progress in recent years,

a really hot area. John Hopcroft told us today that there are still a lot of things we do not know and not understand about machine learning, and that will be our first area, our first cluster of talks. Now, with the young researchers who will be on stage very, very shortly, we just had a three-hour practice during this HLF,

and for them, it's a very special occasion and a very special audience, so you should really give them a big line of applause now, and David Stutz is the first to enter the stage.

He's working at the Max Planck Institute for Informatics, and he will be talking about disentangling, adversarial robustness and generalization of deep neural networks. His hello might sound a little familiar to you, as you've been here in Heidelberg for several days now.

It's on. OK, the hand mic is not on right now. Try again. OK, no, it's working. Magic! OK, good talk. So, my PhD work was actually introduced very well by John Hopcroft earlier.

I'm also interested in understanding the existence and the implications of adversarial examples for deep neural networks, and just as a brief reminder, here we have a deep neural network that correctly classifies this image of a dog as a dog,

but as soon as we add a slight imperceptible perturbation, the dog is misclassified. In this case, it is classified as an ostrich. These adversarial examples demonstrate two important problems with today's machine learning systems. First, there is the security threat that we need to address to kind of deploy these systems

in a lot of applications that we are interested in. We just heard, for example, a lot of medical applications. And second, maybe more interesting, they reveal a serious lack of understanding, especially for these deep neural networks. In the past year, I have been concentrating on the relationship

between robustness against these adversarial examples and generalization of deep neural networks, so their ability to generalize their knowledge to new unseen data. In practice, increasing robustness usually results in decreased generalization performance. And to empirically study this problem, I also assume an underlying data manifold

instead of a cat manifold. Imagine a manifold of handwritten digits here. And then we distinguish between two types of adversarial examples, regular adversarial examples that we show to leave the manifold and adversarial examples constrained to the manifold. We show that these two different types of adversarial examples also induce different notions of robustness.

And based on this distinction, we argue that robustness and generalization are not necessarily contradicting goals. Overall, this means that both robust and accurate models are possible. However, in practice, we still need to find algorithms that do so efficiently. Thanks, and make sure to stop by my poster later.

And the next person on stage is Dragana Raduicic, and she is coming from the Vienna University of Technology, and you see her topic on the slide,

on recurrent neural network for limit order books. Welcome. Thank you. Good afternoon. In my research, using the concepts of supervised and unsupervised learning, we tend to describe the stock market behavior and to analyze the informativeness of the limit order book.

So, the job of the limit order book is to keep track of all unexecuted limit orders. As you can see, the limit order book is defined on a discrete price grid, and each point on the price grid represents the price level.

The orders that need to be bought are placed at the bid side, while the orders that need to be sold are placed at the ask side. The very first step is to dynamically reconstruct the limit order book and to extract the features that are relevant, such as the number of executed trades, open price, close price, and so forth.

Moreover, we enhance our set of features by employing the technical indicators that are standardly used. So, the main idea is to classify each vector of data set

into the one of the labels from the set, idle, sell, or buy. But from such a huge set of features, how to distinguish the features that are relevant and to distinguish from the ones that hold less information.

Therefore, to each feature, we assign a function score with respect to the different measures, namely to the autocorrelation, cross-correlation, and mutual information. Then, we evaluate if the performance of the model which is fed with features extracted with respect to the score function

is better compared to using the randomly chosen features. Thank you very much, Dragana. And now, we will have on stage Bruno Jochen Skrissi.

Please welcome him. He is coming from the Federal University of Rio Grande del Sol in Brazil, but also is associated with the nearby Castro Institute of Technology. Here in Germany. He will be talking about ranking relevances from neural networks for the selection of biological features. Yes. So, how can we better understand

really complex biological phenomena from high-dimensional data? For instance, if you have a few thousand genes, how can we understand which of these genes are related to some kind of cancer

and which are just related with like normality or another kind of cancer? So, there are many ways to do this, but in our approach, we are using neural networks as a map between the features and the conditions. This can be applied for a right variety like cancer,

but also for another kind of biological phenomena. So, here we have a schematic of how it works. We have the biological data come from instrumental laboratories, and we actually train just a simple deep neural network to classify the features into a given condition.

And we then do the interpretability methods, such as layer-wise relevance propagation, to interpret the results of the neural network. So, we get the output of the neural network and find out what are the relevance of the individual input to this output.

Then we run this relevance for each sample of the data, and then we obtain a table like this one that gives us the relevance of each, for example, gene, the relevance to classify the data in the cancer or health tissue, and also we can explore this data in the class level,

so we can aggregate over all samples that have cancer and not have cancer, or globally, so we can better understand this data overall. So, this is early stage work, but I would be happy to discuss the preliminary results by my poster. Thank you very much.

Thank you, Bruno. We will now continue with Haminshu Yan, who comes from the Indian Institute of Technology and will talk about extreme classification. Welcome.

Namaste. My research topic is extreme classification, which is basically multi-label classification with millions of labels. So, traditionally, people have been solving multi-label problems with, let's say, hundreds of thousands of labels, but we are looking at problems with millions of labels. So, the objective is to annotate data points

with the most relevant subset of labels from an extremely large label set, where extremely large is like millions of labels. For example, if you look at Wikipedia, each Wikipedia page is tagged with certain categories. So, overall, there are about a million categories. Now, you want to build a classifier such that whenever a novel web page comes in, you can automatically tag it with the most relevant subset of categories

out of these million categories. So, what are the possible applications? So, in general, any recommendation problem can be reformulated as an extreme classification problem by treating each item to be recommended as a separate label. For example, if you look at this related searches problem, where the objective is that given a query that is issued on a search engine,

the objective is to recommend related queries that might solve users' information requirements better or might provide more information on that same topic. For example, if you search for Heidelberg Laureate Forum on Google, some of the related searches are Heidelberg Laureate Forum acceptance rate

or Heidelberg Laureate Forum wiki, et cetera. So, this can be reformulated as an extreme classification problem by treating all these related queries as separate labels. So, we developed an algorithm called Slice that can scale to as many as 100 million labels, and we used it to solve the related searches problem.

So, the algorithm gives state-of-the-art accuracy and has long-time training as well as prediction, and we deployed it on the Bing search engine, and it improves not just the coverage of the queries that it can serve, but the success rate in terms of click rate as well.

So, if you are interested, come to my poster. Thank you. Thank you. And next, we will have Amanda Bauer from the University of Michigan on stage. Please welcome her, and she will be talking about preference modeling

with context-dependent salient features. The floor is yours. Hi, good afternoon. Part of my research deals with learning a ranking on a set of items given human comparison judgments. This has applications in recommender systems and understanding human perception. My work in particular aims to model human judgments in context.

So, for a concrete example, suppose our set of items are US legislative districts, and we wish to rank these districts with respect to compactness. Typically, what's done is pairwise comparisons of the form District A is more compact than District B are collected,

and then aggregated to learn a global ranking. We do this because it would be difficult to ask each of you to rank thousands of districts at once. However, what's implicit in this method is that the mechanisms by which humans make judgments on a global ranking scale are the same mechanisms by which humans make judgments when they make pairwise comparisons.

However, this kind of approach doesn't allow it to model intransitivities in preference data, which does arise in some data sets. So, my work aims to model the way that we make judgments on a global ranking scale different from how we make them in a pairwise comparison setting. So, for example, if I asked you to compare the first two districts

with respect to compactness, something like convexity as a feature pops out as salient, and you might only use that feature to make the comparison. However, if I asked you to compare the second and third districts, they're kind of similar on all dimensions, and you might use all the features.

So, we have a probabilistic model for saying the probability that one item is better than the other in context, and we can show that with enough samples, maximum likelihood estimation can accurately estimate the underlying ranking. Thank you, Amanda.

And we will move on quickly with Ramana Rayan Mohanty from the Indian Institute of Technology, and his topic is graph-based machine learning techniques for classification of hyperspectral images. Please go ahead.

Namaste. So, first of all, I will start what is hyperspectral imaging. So, hyperspectral imaging is a special kind of image that is mostly captured by the remote sensing satellites or UAVs, and that image has a hundred number of bands. That is the major challenge, and that is the major motivation

that motivate us to go for this research. So, out of my PhD research, I have figured out two major problems here. So, the first problem is the data geometry distribution problem, and the second problem is the class discrimination problem. So, the data geometry distribution problem is nothing but...

So, while we take in hyperspectral data, mostly the data is non-linear in structure, and the majority of the algorithms, they just ignore this non-linearity property, and they simply flatten the data and go for any classification task, so that once they flatten the data, the data geometry got distorted, and once the data geometry got distorted,

the classification performance got degraded. Similarly, for the data, in the class discrimination problem, the majority of the time, what happens, the similar type, different classes have similar type of features or similar spectral signatures, so that it is quite difficult to figure it out which class

and figure it out the classification accuracy, and also it affects the classification accuracy a lot. So, to solve these problems, I have devised two different approaches. So, in the first approach for this data distribution, data geometry distortion problem, I simply went for a manifold mapping approach. In the manifold mapping approach, I mapped the original data,

non-linear geometry, to a new discriminative geometry, and there, I just simply go for flattening and then classifications, so that it doesn't disturb the data geometry and it will give better results. Similarly, for the class discrimination problem, I went for a spatial spectral approach,

and with non-linear data, it gives us the same supervised approximation. So, this is the results for this thing. For your interest, if you are interested, please hop on to my postdoc. Thank you very much. Thank you.

And next will be George Boa-Teng from the ETH, ETH Zurich, and it was a surprising topic for me, multimodal emotion recognition among couples for diabetes management.

Akwaaba. My name is George Boa-Teng. I'm from ETH Zurich, and I'm from Ghana. So, evidence suggests that for married adults, illness management is shared with spouses, and it involves social support.

Social support has been shown to result in healthier habits among diabetes patients, and also have either positive or negative effect on emotions. Hence, through emotion recognition, we could have a real-time assessment of a key outcome of social support, which is emotions, and through that, we could develop various interventions

to help couples better manage chronic disease. So, in this work, our goal is to address that gap, and in collaboration with health psychologists, we've developed Diamond, a mobile and wearable system that collects multimodal sensor data from smartwatches. So, we are collecting data such as audio accelerometer, gyroscope, heart rate,

and we're collecting data when the couple is close and when they're speaking. So, to detect closeness, we're using the Bluetooth signal strength of the two smartwatches, and to detect if you're speaking, we developed a voice activity detection model, which runs on the smartwatch.

And then we also collect data about emotions, and then the plan is to use that information to develop an emotion recognition model using various machine learning and deep learning models. And so, there are three potential contributions of my research work. The first is some kind of novel mobile and wearable system

that is able to assess how couples manage chronic disease together. The second is some kind of machine learning model that is able to do emotion recognition using multimodal sensor data from smartwatches from the real world. And then, finally, another contribution would be

some kind of a novel wearable system that's able to recognize the emotions of couples. If you want to learn more, come to my poster. Thank you. Thank you. Thank you. Thank you. OK, the next person on stage will be Gadir Abouda,

and she's coming from the Hamad Bin Khalifa University in Qatar, and we'll talk about link prediction via higher-order motive features. Please welcome Gadir.

In this session, my thesis of the project, it's about link prediction. Yeah, so, link prediction is a common problem in the graph analytics, where the goal is to predict the new edges that will be added in the graph in the future

that connect unconnected nodes. An example of such a problem is predicting new friends in social networks and also recommending versions of, like, products on such a graph. In my research, I did an analysis on the deepest structure

that's around two nodes in order to study their connectivity by extracting some big graphs that surround the two nodes, or the pair of nodes appears in, and use that as a feature for my classification model.

So, for example, we have here different six shapes that we can get extracted from four nodes, and we did the analysis on different types of... different number of nodes in order to use this feature for the classification, which can be generalized for many types of the graph, and also to improve the accuracy.

As we can see here, it's compared to many, like, state-of-the-art methods that are used for the link prediction and using that analyzes the typical features at a certain level, and also compared to many complex models that are used for this analysis.

And we did this conversion and training dataset that we created also in this project. If you want to know more about my research, come and visit my poster. Thank you. Okay, now it's time to change to another cluster,

and we will be in the field of analysis right now, and already on the stage is Jan Glaubitz, who is working at the Technical University Braunschweig and also is associated with the Max Planck Institutes for Mathematics. He'll be talking about numerical integration of experimental data.

Go ahead. Okay, thank you, and guten tag, Madananda. So, within the next minutes, I want to get you excited about numerical integration. And to demonstrate that you have very good reason to be excited about numerical integration, I want to start with an example, addressing a hot topic, which is climate change.

So, imagine we have a power plant, and imagine we want to calculate the total amount of carbon dioxide emissions of this power plant over a time period of one year. So, what we could do is, we could jump into a plane, fly over the power plant and do some measurements. And we could do this, let's say, once a month or once a week.

So, what we will end up with is a finite set of discrete measurements, illustrated in the lower right corner. The approach then is to approximate the total amount of emissions corresponding to an integral by a finite sum over our weighted measurements. And such a sum is called a quadrature rule.

Now, there's only one thing left for us to do, which is to determine rates v, which give us a good approximation. Usually, people tend to use one of two approaches. In the first approach, they optimize for the degree of exactness, which means that polynomials up to a certain algebraic degree

should be treated exactly by our quadrature rule. This yields to the class of inter-policatory quadrature rules. Unfortunately, these are well known to be unstable for equidistant measurements. So, in the second approach, one therefore goes for composite quadrature rules. These are known to be stable, yet of limited degree of exactness,

and both are illustrated on the lower right corner by the black and red line. So, on the one hand, we have this optimal degrees of exactness, and on the other hand, we have stability. But I am greedy, and I want both. And in my recent research, I have been able to get both. And the essential idea was to aim for something between,

to aim for a high degree of exactness, but not the optimal one, and then to use the remaining degrees of freedom to also ensure stability. So, if you're interested in hearing more about this, feel very free to visit me at my poster. Thank you. Okay. Thank you.

Now, we've got Shin edu. He's a Shukwu from the University of KwaZa, Zulu, Natal in South Africa, and he will be talking about iterative algorithms for a finite family of monotone inclusion problems in Hadamard spaces. The floor is yours.

Thank you for the introduction. Dalunu, my research project is on the development of iterative algorithms for solving monotone inclusion problems. Now, the question is, what is monotone inclusion problem? As defined in the problem definition above,

we have that monotone inclusion problem has to do with optimising industrial problems, for example, minimising the cost of production in an industry. But this problem has been studied

by so many researchers in Hilbert and Banach spaces. But the problem here is that these spaces, where this research has been conducted, has a lot of limitations when it comes to application. So, because of this,

this motivated us to extend the research to a more generous space, for example, the Hadamard space, where we can actually have some nice applications. Besides the application of minimising the cost of production in an industry,

we have other nice applications in Hadamard spaces, such as the shape analysis of three light structures, computing the averages of four threes, and again, modelling of airways of systems in human lung and blood vessels. This we intend to go into deeper,

because our research so far has shown that solving this problem has applications in the above listed. Now, we also intend to go deeper in the last one, the modelling of airway systems in human lungs and blood vessels.

We actually intend to do some work in optimal control problems. Now, this is the model we have, but because of time, if you are interested, I can explain for you the iterations and the simulations. Thank you.

Thank you, Sunidu. OK, things change quickly here, and we will move on to the next cluster, which will deal with computer architecture, and we've got Rahul Jain from the Indian Institute of Technology in Delhi.

Welcome, Rahul. And he will talk about machine-learned machine reinforcement learning exploration for architecture co-optimisation. Yeah. Namaste, Heidelberg. This is Rahul from India.

So the topic is machine-learned machines. That's my PhD topic. And what we explored in this is, can we use machine learning to improve the machines themselves? So at the top level, the research problem is really effectively applying multiple optimisations on a microprocessor at runtime,

and we have used multi-agent reinforcement learning to do that. The main contributions are we have tried to apply this technique to two co-optimisation problems, and we have explored three different multi-agent architectures, namely independent multi-agent architecture, wherein the agents are completely independent,

they're not aware of any other agent on the system. Cooperative, wherein these agents are still independent, but share information for better decisions, and coordinated, wherein they actually jointly work together to perform a global optimisation. The diagram, the figure shows how a typical resource allocation would really work, which we could discuss over the poster.

The first co-optimisation problem is DVFS and DCP, wherein the DVFS is an energy-saving technique which can result in some performance degradation, and dynamic cache partitioning is a technique to improve the performance. So we have explored how we can apply both of these optimisations together and achieved almost more than 20% energy delay product improvement.

The other technique is dynamic cache co-partitioning, wherein this was a new problem which we proposed and showed how we can perform cache partitioning over multiple cache levels, and we were able to show almost like 9% system improvement. We got a few recognitions for the thesis,

the most notable one being that the thesis was recognised as the best PhD thesis at the 32nd International VLSID Conference, and Intel has filed for a patent on some of the work. Thank you. Thank you.

And next, we've got someone from Princeton University, Yatin Manarkar, and he will talk about automated formal verification of event orderings in parallel systems. Go ahead. Afternoon all. When building systems, we tend to think of verification as something that happens pretty late in the development process,

so after implementation. But this means that bugs end up being found pretty late too, and some of these bugs might be issues with the design itself, necessitating a redesign and resulting in the loss of development time spent creating the incorrect implementation.

Formal verification can provide strong correctness guarantees based on mathematical proofs. I propose continuous formal verification, or formal verification throughout the design process. In this philosophy, formal models of systems are created and verified at early stage design, thus catching design bugs before implementation commences.

These formal models can evolve with the design, and the eventual implementation can then be verified against a detailed formal model to help ensure correctness. My dissertation has enabled continuous automated formal verification

of memory consistency model properties in parallel systems. Memory consistency models, or MCMs, specify the ordering rules for memory operations that are used for synchronization in parallel systems, so MCM verification is critical to parallel system correctness. I've developed automated tools and methodologies

for formal verification of MCM properties in early stage designs, detailed designs, and Verilog processor implementations. The automation of the tools makes them easy to use, while the formal nature of the verification provides strong correctness guarantees.

If you'd like to know more, please come by my poster. Thanks. Thank you. And now we've got Abdul Rahman Mahoud from the University of Illinois at Urbana-Champaign. The floor is yours, and he will be talking about the Minotaur,

or about adapting software testing techniques for hardware errors. Hello, everyone. Ahlan wa-sahlan. So today I want to talk about one of my projects, which is called Minotaur. And the motivation for this project is that transient hardware errors

are becoming ubiquitous and an important problem to solve in many systems. For example, a single bit flip in an application, in safety-critical applications such as self-driving cars or medical devices, can have catastrophic events down the road. So traditional hardware redundancy techniques are too expensive,

which usually require doing redundancy either in time or space, so throwing two pieces of hardware to compute or running the same application twice. A lot of recent techniques looked at software-directed approaches which tried to find anomalous behavior at the software level to detect hardware errors.

So overall, these have been pretty good, but one problem is that a lot of these errors, or some of these errors escape as silent data corruptions. My research focuses on resiliency analysis, which is analysis to try to find these silent data corruptions and protect against them before deploying a system. One problem is that most of these analyses are very slow,

or they're either very fast or very accurate, but not often they're both. My work, which I coined Minotaur, tries to do both. The key insight behind Minotaur is that analyzing software to find hardware errors is similar to analyzing software to find software bugs.

My research tries to bridge the gap between these two fields by adopting four software testing techniques which are shown on my right, to hardware resiliency analysis in different forms. By doing so, we showed that we could get an order of magnitude speed-up and much higher accuracy compared to the state of the art. For more details, please come by my poster and talk to me more.

My poster is in the far back corner, so please don't miss me. Thank you. Okay, now it's time to do some algebra as we go to our next category, and we have here Ram Lopez on stage from Cleveland State University, who will talk about algebraic methods for coding theory.

The floor is yours. Hola. So, I studied in my dissertation coding theory using commutative algebra tools. But what is the meaning of coding theory? Well, coding theory is behind any electronic device

every time that we want reliable, transmit information. For instance, if you are using your cell phone or your laptop to send an email, you are using coding theory. The idea of coding theory can be seen from this very simple example.

It's clear that there are some errors, that there are some mistakes, but even in that case, we are able to read and understand every single word, and one of the reasons is because of the context. So that is what coding theory does.

It creates the context in such a way that it's going to be able to find and correct the errors. But they want to be clear that coding theory is not only to correct text. It is much more than that. For instance, if you are using your cell phone to read a QR code,

you are using also coding theory. What I studied in my dissertation is coding theory using commutative algebra tools. Coding theory is a branch from information theory. Commutative algebra is a branch from pure mathematics. Would you want to know more about the bridge between information theory

and pure mathematics, please come to my poster. Thank you. And next in line, we do have Mahendra Gruptra from the National Institute of Technology,

and he will talk about the necessary and sufficient condition for existence of observer for the scriptive system. Please, go ahead, Mahendra. Namaste. So there are two terms in the title of my talk. One is descriptor system, one is observer.

So I will go them one by one. So the system one is the mathematical model for descriptor system. In this system, everything is known except the x. So x is the state which is unknown. So x is to be estimated with the help of one more mathematical model. We call it observer.

Now fact says that there exists an observer for system one if and only if system is causally detectable. So causally detectable is the combination of two conditions. One is the detectable that the original system is detectable, or the second one is the extended matrix triple is causal. So my problem is to identify the causal detectability

in terms of system coefficient matrices. So we have solved this problem in separate parts. So first I have established a causality condition directly on system matrices. Then I have extended this condition for the extended matrix triple.

So that is given by equation three. Now this equation three is combined with the condition for detectability, that is equation four. Now finally as a result we can say that there exists an observer for system one if and only if equation three and four are satisfied. So for all these theorems, I have given proofs in my poster.

Thank you. And now we will proceed with Niraj Sanghwan from the Indian Institute of Technology in Bombay. And he will talk about integrally closed simple extensions of evaluation rings.

Namaste. My presentation is about integrally closed simple extensions of evaluation rings. Mainly the work is about algebraic number fields and when we can say that the product of rings of algebraic integers of two algebraic number fields is integrally closed.

So for the general things, let Rv be the value of the integral value of v of a field k. And in 2006, Ayrshow gave a criterion for when Rv theta is integrally closed for an element theta integral over Rv. Using this criterion, we were able to give explicit conditions

based on the coefficients of polynomials of a trinomial to when this Rv theta is integrally closed. And using that result, we were able to prove this result for quadratic fields, that when k is an algebraic number field and L is a quadratic field not contained in k,

then a, k, a, l are closed if and only if discriminants of k and L are co-prime. Another motivation for this problem comes from this,

that if k1, k2 are two algebraic number fields linearly disjoint, then if their discriminants are co-prime, then they are linearly disjoint. And another thing, that the product is integrally closed. This product, and using this, we proved a general result for valuation rings.

And through that result, we were able to obtain this final result, that if a, k1, a, k2 is integrally closed, then the relative discriminants of k1 and k2 are co-prime. Thank you.

Thank you. Now it's time to listen to another European language, as we have Idin Lidan on stage, and he is coming from Bosnia and he's a governor from the University of Biha, and we'll talk about topological characteristics of tilings on surfaces.

Hello, everybody. Firstly, I want to thank Heidelberg-Laureate Forum for enabling me to be with you today. Today, I will be going to talk about the topological characteristics of tilings on surface with polyominoes.

This is joint research with my PhD advisor, Giorgia Baralic. Goals in our research are studying the problem of tilings on surface from topological combinatorics and algebraic aspects. Give a proof that non-triviality of the obstruction element in the homology group of tilings and existence of tilings.

Give some generalisation for some class of problem in higher dimension. Our work is continuation of research of Kornwey and Lagarias in their remarkable paper, Tiling with Polyominoes and Combinatorial Group Theory, and Ray's strategy with working in finite homotopy group

in his paper, Tile Homotopy Groups. We have developed their idea on context of topological surface in our paper, the homology of polyomino tilings on surface. Now, I will present the method which we used.

In topology sense, we consider a topological way for creating the surface from a polygonal model of surface, gluing its sides. In combinatorial sense, we consider, analyse a set of polyomino shapes

and grid the property. And in algebra sense, we are working with the quotient group to calculate the homology group of tilings. The obtained results are presented on my poster on the hall for the posting session, so welcome, thank you.

Thank you. And now, the next cluster is distributed computing, and we will have listened to Suman Saraf from the National University of Singapore with slow links, fast links, and the cost of gossip.

Hello, namaskar. So, I'm going to talk about the slow links, fast links, and cost of gossip. Consider that all of us form a network here... Sorry. Consider that we all form a network here,

and I am this evil red node that I want to spread a rumour among all of us. So, none of us can come and announce what the rumour is, but what we can do is we can go up to someone, go up to his ear and say, pish, pish, pish, pish, this is the rumour. So, in each round, everyone who knows the rumour

goes up to someone and shares the rumour. Everyone who doesn't know the rumour asks someone what the rumour is. So, the end goal is to let everyone know what the rumour is. So, this is how the communication goes. OK, OK.

Sorry. Oops. Yeah. So, if everyone could talk to everyone, this rumour spreading is really fast, but that's not always the case. Right. Suppose I restrict, like, only my friend here from the third row, could talk to my friend here in the third row, and none of these people from this side could talk to the people from the other side.

Then it takes longer for the rumour to spread, like, say, in this scenario. So, what characterises how fast we can spread the rumour? So, the thing that characterises the fastness that the rumour can be spread is called conductance of the graph. But here, in this case, what we did is we ignored that each communication takes the same amount of time.

So, normally, if you were to say a rumour to a friend close by, it will take you less time. But if you were to choose a person, someone far away in the other end of the hall, it will take you longer. So, what happens when these edges here have latencies? Is this graph faster? Or can I spread a rumour faster in this graph?

Or in this graph? So, we tried to answer this question in my walk. So, what we do is we define a parameter called critical conductance, and we show this characterises the speed of communication in networks with latencies. To know more, come please visit my poster.

Thank you. Can't pick quicker. OK, now we've got Caleb Stanford from the University of Pennsylvania, and he will be talking about automated code distribution for distributed stream processing.

Thank you. So, in the future, a large amount of data will be generated by Internet of Things devices. Such devices include, for example, medical devices that are implanted,

live video cameras and drones, as well as mobile phones, of course, as we saw in the last session. So, in all these examples, you have a large number of nodes where distributed data streams are arriving, and these data streams are arriving at an extremely high rate, and your goal is to make a decision in real time

at one of these nodes in response to all of this large amount of data that's arriving. So, my work is about specifically making programming for such applications much easier. So, the end-to-end system is as follows. You would have the programmer who writes a specification,

and then the specification is sent through the system through a compiler, through an optimizer, and through a code distributor, which eventually produces code at all the individual nodes in your distributed system at all the different devices. And why does this make it easier? Well, first of all, we want the specification to be high-level.

So, you're not worrying about the low-level details, you're writing a high-level specification, and then the system distributes the code automatically. Second, we want to optimize automatically. So, rather than worrying about the performance at a low level and trying to manually optimize your code, the system will take in an arbitrary performance metric

that you give it and try to optimize for that metric. And finally, because the code is distributed automatically, you don't have to worry about concurrency. So, concurrency is a major problem when you're talking about correctness of these applications, and here, you write the specification without worrying about concurrency, and we preserve the correctness when we do the distribution.

So, for more information, please come see my poster. OK, now please welcome Antonino Galletta from the University of Messina in Italy, and he's already perfectly dressed for the Bavarian evening,

though you might not recognize that because it's the perfect Italian suit. Go ahead with innovative cloud- and edge-based systems for smart life. Thank you for presenting me. I'm Antonino, and today I will present to you my research activities.

They will be divided into two main branches, smart cities and health. In particular, with regard to this latter, we are working with a research center in Messina that provided us three research questions. So, how we can share medical resonance images with foreign physicians in a secure manner,

how we can analyze and manage the huge amount of data that our medical devices are producing, and finally, how we deal with general data protection rules. So, in order to solve these questions, we developed this architecture. It is composed by several microservices,

each one with a specific function, like us, the anonymizer, that you can see in the middle. Okay, so it is in the middle. And that allows us to work with GDPR. Instead, on the left part, you can see microservices that we developed in order to work with big data. In particular, we developed some big data visualization tools.

Finally, on the right part, you can see how we are sharing MRI among clinical institutes. We are using public clouds, such as Google Drive, Amazon, and the secret share algorithm. By the way, for future details, come in my poster.

I will wait for you in front of room H2011. Bye-bye. Ciao. Thank you. Now, we've got a lot of bright minds here, so it's no problem that we change the field once again and move on with differential equations.

And we will start with Analia Silva. And Analia Silva is coming from the Institute of Applied Mathematics in San Luis in Argentina, and will talk about elliptic problems with non-standard growth and lack of compactness. Thank you.

Hola. So, this is not mine. Now it should work. Just try. Other side, the other button, then you're fine. This is mine. So, Solen Embedding Theorem says that Solen Embedding Space

is embedding in LQ, the embedding is continuous if... Oops, sorry. If... Oh, you're good. Sorry. Just turn it the other way around, probably. OK, sorry. If... Sorry. If Q is smaller than the critical exponent P star,

and moreover, we can define the constant S as an infimum of this quotient, and the constant S is called the best constant in Solen Embedding. The theorem says that this constant is positive, and if there exists B that satisfies this,

it's called an extremal. So, the theorem says that the embedding is compact. See, if Q is strictly smaller than the critical exponent. So, for the constant case, the key point is the concentration component seen principle by Ljons. So, for the constant case, the best constant is not attaining.

The constant doesn't depend on the domain, and the Euler-Lagrange equation has no solution and any start sharper. And we deal with the critical... the critical problem for variable exponent space and what results are.

We extend CCP in the variable exponent setting, we prove existence of extremals, we prove existence of solutions, and we obtain the same results for the sublet tracing embedding. So, to know more about this, please visit my poster.

Thank you. OK, and now we're going to listen to Mavish Ali, who is coming from the Yamea Milia Islamia New Delhi, India, and she will talk about integral transforms

and generalized form of hybrid Bessel functions. The floor is yours. Hello, everyone, and aslamo alaykum. My work is on introducing new families of spatial functions and to find their properties. In this work, we have introduced the one-parameter extension

of the Laguerre-Gould-Hopper-Bessel functions, which is given by equation one and introduced by Hahn et al. in their paper published in Reports on Mathematical Physics. So, our main objective is to find their generating function, their explicit representation and the integral representation.

The approach we used here is the use of Euler's integral and certain combination of operational rules to obtain the generalized form of hybrid Bessel functions, which are defined by integral transform three. And using this integral transform three,

we obtained our desired results. These are the results. The first one is the generating function of the generalized hybrid Bessel functions. The second one is the explicit representation. And the third one is the integral representation of the generalized hybrid Bessel functions. To know more, come to my poster.

Thank you very much. Thank you. And now we've got a very long title, Boundary Value Problems for Third-Order Linear Loaded Differential and Integral Differential Equations, mixed type. And we will listen to Omidah Batayeva,

who comes from Uzbekistan from the Khorsum Mamu Academy. Go ahead. Thank you. My name, my work about boundary value problem for the mixed type equation. On the other side, it's the equation of the loaded.

The theory of the mixed type equation, it's one of the principle part of general theory of partial differential equations. The interest for this kind of equation arises both theoretical and practical uses of the application. For example, the mixed type equation arises in Gaussian dynamics, magnetic heterodynamics, and other and other.

On the other side, in the recent years, with intensively research on the problem of optimal control agroeconomical systems and a lot of problems, mathematical physics and mathematical biology, for example, in the problem of dynamical population

and some of the investigation problems, it has become necessary to investigate new class equation. It's called as loaded equations. The equation, equation one is called loaded if the operator R depends on the restraction of the unknown solutions,

belonging in the closed subset D. And very important case is equation one if it's possible to write in the form two, where L is the differential operator, M is the loaded part, which can include combination of the unknown solutions,

the trace of the unknown solution with the integral differential operator. In our investigation, we are investigating the main boundary value problems for the mixed type loaded equation as the problem of trichomy, Heller's steps, Darbock's and the generalization, and also well-posed new boundary value problems

for the mixed type loaded partial differential equation of the second and third others. Thank you for your attention. Okay, let's get the slide straight, then it looks nicer in the picture.

And we will now listen to Bristol, to the University of Bristol, Macrina Agaroklu will come on stage, and she will talk about bifurcation and stability analysis of periodic solutions in nonlinear lattices. Here we have a simple first order equation.

It comes from the Kirchhoff voltage law for this electric circuit. But what is the connection between this electric circuit and Harry Potter's invisibility clock? Researchers find out that they can create a kind of invisibility device using metamaterials.

Metamaterials are materials that we cannot find in nature, but we create them in the laboratory and they have the properties that we want them to have. For example, negative reflection index. And usually the concept of metamaterials refers to a periodic arrangement of artificial particles called split-ring resonators, as we can see in the figure on the right.

And each split-ring resonator can be modeled as an electric circuit. The last years, many companies used metamaterials to build their products. As a part of my PhD thesis was to solve this kind of equations, but more complicated with coupling and with nonlinear terms

using Melikoff method and numerical simulations. We studied the existence and persistence of travelling waves and the stability of the solutions. Of course, as I said before, there are multiple applications like for satellite antennas, security industry, radar for drones and magneto-inductive lenses. Thank you very much.

And our next speaker will be Christian Offen, who is working at Massey University in New Zealand and he will talk about structure-preserving numerical integration.

So, I would like to talk to you what my research in geometric integration is about. It's all about making errors, but not mistakes. If you calculate the solutions to an ODE or a PDE on a computer numerically, there will be approximation errors, there's no way around.

The trick is to make sure that these errors do not change the qualitative behavior of your solution. So, which kind of qualitative aspects am I looking at? I'm looking at bifurcation problems. So, here's an example. It's a boundary value problem for an ODE with one parameter.

It has, if the parameter is between zero and three, it has two solutions, a high energy solution, a low energy solution, and as you increase your parameter, these two solutions will merge and annihilate each other. As you see here, if I plot the parameter versus the energy.

Now, that's a very simple bifurcation diagram. There are more complicated ones. For example, an elliptic umbilical. Here we go. So, there it is. I've proven that there are a lot of bifurcations which are related

to something that's called symplectic structure. Now, if you preserve the symplectic structure in your numerical computation, you will make errors, but you will see the correct bifurcation diagram. If you destroy the symplectic structure, you'll see the wrong picture.

Now, researchers calculate these bifurcation diagrams to understand the model, and we would really like them to conclude from the correct pictures. And this is what my research is about, to make sure that approximation errors

don't lead to mistakes. Thank you. Thank you. Okay. Now it's time for our last category, and the organizers found it a little difficult to put these presentations you're up to listen to

into an exact category. So, this is called miscellaneous, which is sometimes the most interesting part in a newspaper. So, we are really surprised and wondering what lies ahead of us, and we will start with Shwadari Ahmed from the Singapore University of Technology and Design,

and he will talk about noise print in extrusion detection in industrial control systems through sensor fingerprinting. Please, go ahead. Thank you, and Khusham Deed. So, my research focuses on cyber-physical systems, especially the security issues. So, I should give a few examples before starting.

So, good examples are smart grid, water treatment plants, and autonomous vehicles. One important thing about these cyber-physical systems is they have a lot of sensors, and these sensors measure the physical quantity, and they send the quantity information to the controllers, and they should take some action based on the sensor measurements.

And the problem we are trying to deal with is how are we sure that the readings the sensors are sending to the controllers are really coming from our own sensors, or have been modified in some way while communicating to the controllers, or have been distorted in the analog or physical space.

And towards that end, what we are trying to do is do sensor data authentication, but without using cryptography for several reasons. One is that sensors are not computationally that powerful. And so we came up with an idea called noise print,

and essentially what we are trying to do is create a fingerprint for the sensor based on the noise of the sensor. So we agreed that whatever good sensor we have, it does have some component of noise in the measurements. So we do a couple of hypotheses. Can we distinguish sensor based on the noise profile of the sensor?

And secondly, can we distinguish between same type of sensors? To give a very quick example, imagine that you have this sensor, which is making a constant distance from the wall, and we get the distance, yeah, but it is not free of noise. And this is the whole idea. Can we do sensor fingerprinting based on the noise measurements?

So we have some results. On the left, what you see is time domain analysis, and we see that we have same type of sensors, and we are able to distinguish between them based on the statistical features from the noise. And on the right, we see the frequency analysis. So we have done this on real systems, on hundreds of sensors, and it works pretty well.

For more information, please look for our paper called Noise Print. Thank you so much. Thank you. Okay, please welcome Adaranki Zakpere from the University of Ibadan in Nigeria, and he will talk about evaluation of three privacy,

or she rather, sorry, will talk about three topics. Evaluation of three privacy schemes, K, anonymity, L, diversity, and T, closeness. The floor is yours.

Thank you. All right. So this research work was carried out when I was a PhD student at the University of Cape Town, and I was then sponsored by Astoplatna Institute at Boston. All right, to just give a motivation for why we embarked on this project,

we discovered, especially in Africa, that most of the time when we need data for research is always very difficult. It's difficult to get the organization to release data to researchers or third party, mainly due to trust issues. And we were interested in working in the crime space

in South Africa, but we also had difficulty getting ourselves to data, to crime data. So we thought of why can't we create an intervention that we anonymize data, such that when people, when third party need this data, they can make use of this intervention, and the data will be anonymized and released

for researchers or third party to make use of. And that led us to this question, privacy versus data utility. How do we balance privacy and data utility? How do we anonymize data in such a way that it will still be useful, and at the same time, the privacy level with the maxima?

And that led us to literature review, and then we decided to make use of three different types of privacy schemes called k-anonymity, l-diversity, and t-closeness. And these anonymity schemes, they work in very simple way. The three schemes make use of generalization.

That means when you have a crowd of people, the privacy scheme tends to cluster, tends to apply clustering technique on those group of data, and through that, privacy can be carried out. And yes, this is one of the results that we got from the experiments we ran,

and we realized that as we increased privacy level, then that type of utility becomes minimum. Thank you very much. If you are interested, please come to my poster. Thank you. Thank you.

We will stay in Nigeria, but switch universities, and the next speaker is working at Lagos State University. It's Kayode Oshinubi. Please welcome him on stage, and he will give a presentation on vibration analysis of non-uniform Timoshenko beam

resting on elastic foundation. Yeah, thank you. Eka-san, the motivation of this work is to model an equation, a partial differential equation that we are gonna be using, the Timoshenko beam theory,

and majorly want to look at the foundation parameter, which is a two-parameter foundation, elastic foundation can be one, two, three parameter foundation, and vibration analysis makes us to look at the frequency,

the natural frequency, and the mode shapes of how the model equation behave. My equation is very complex, but if you can visit the poster, I'm going to explain more about it, and the equation is complex because we are dealing with non-uniform model,

so that is why you can see some of those parameters there, and the approach we want to use is, the first, we'll put on a zoom solution. From there, we'll non-dimensionalize it, and after non-dimensionalizing, we'll solve using Adamia

and differential transform method, and also we'll apply some compatibility conditions and some boundary conditions, which we didn't show you, but if you can visit the poster, we'll show you, and we'll also draw the mode shapes with different natural frequencies to see

how the three different model equations that we're able to put up, yes. Thank you. Time for our last poster. Harini Havwarachchi from Monash University,

Melbourne, Australia, who will talk about exoton plasma nanohybrids for minimally invasive tumor imaging. Great to have you here. Thank you very much. Suppose you take a substance,

a chunk of gold, for example, and crush it down into very small pieces that are in the nanometer scale. If you do that, the resulting substance would no longer have the same photophysical properties as the bulk material. In our example of gold, the particles that result would no longer be gold in color.

In fact, their color would depend on their size. So small particles like this that are less than about 100 nanometers in size are called nanoparticles, and these are so tiny that they may not obey the principles of classical physics that we see in our everyday lives.

Metal nanoparticles and quantum emitters are two such types of nanoparticles. If we shine a coherent beam of light on these particles, excitations known as plasmons form in metal nanoparticles, and another type of excitations known as excitons

form in quantum emitters. And if we combine these two types together, we can form nanohybrids that exhibit fascinating optical phenomena with a range of potential applications, starting from photovoltaics and light harvesting to nanomedicine.

In my research, I have theoretically demonstrated some prospects of using these nanohybrids as feedback probes to reduce the heat-induced damage to the surrounding healthy tissue when photothermal cancer therapy procedures are conducted. I invite you to have a look at my poster

if you are interested. Thank you very much. Thank you. Yeah, didn't they all do a really great job here on stage? Now, it's time to say goodbye in my mother language,

and as I come from the northern part of Germany, where people speak lower Saxon or an East Frisian dialect of that, there's a wonderful expression for saying goodbye, and it's, höldi föchtig. And höldi föchtig does mean, have a good time, stay healthy, and be brave.

Therefore, höldi föchtig, and don't forget to meet all of them at their posters and all of our presenters from the poster flash on stage now for our last final photo. Goodbye.