Usage of deep neural networks is often connected to powerful (multi)GPU machines. However, we want to use their potential on every piece of hardware, even on mobile phones. It is still not straightforward to use them on embedded devices because of their size, resource consumption, performance and the like.
Nevertheless, there are techniques which can be used to make deep neural networks and embedded devices more compatible. We will talk about specialized neural networks architectures, quantization, mobile platform optimized NN libraries and even more.
is a computer vision team leader at Innovatrics
. His team works on problems related to face processing (detection, tracking, recognition, etc.) on various platforms ranging from GPU servers to ARM devices. He received his Ph.D. from FEI STU Bratislava.
Using object detection with CNTK to classify hotel images (Karol Żak)
In our next talk you can look forward to a presentation that will reveal ML application in software products – our speaker Karol Żak will show you how they used object detection with FasterRCNN using CNTK to classify hotel images. All the learnings they share, come from cooperation with customer on real-life solution – platform Hotailor with nearly a million hotels inventory using smart analytics to help businesses sell more hotel rooms.
Karol Żak is Technical Evangelist @Microsoft – a technology enthusiast for years associated with Microsoft and IT communities in Poland. He is also former Microsoft Student Partner, founder and ex-leader of .NET Group in Technical University of Radom. Karol is all about new technologies and trends in IT but mostly interested in cloud computing, AI/ML and mobile tech.
For 2 years as a certified trainer with MCT title, he led IT trainings for developers and then in 2014 joined Microsoft as a Technical Evangelist. Privately he is a big american football fan and starting quarterback for one of Warsaws teams and also Polish National Team.
Implications of adversarial environment on machine learning (Michal Nánási)
Spam is an adversarial problem. Adversaries at the other end of the internet connection are trying to get around your detection. Each time you deploy an anti-spam measure, adversaries are no longer making money. If your service is big enough they are motivated to go around your detection every time you improve it. More sophisticated adversaries can make your classifier worthless in less than 15 minutes. That is one of the reasons why you still see spam today.
In this talk, the we will talk about the problems you may encounter if you try to apply machine learning in adversarial environment; when an adversary is constantly trying to circumvent your detection. Talk will be focused less on specific models, and more on how this environment affects your workflow and the way you approach machine learning problems.
After finishing his PhD in Bioinformatics (FMFI UK, Bratislava), Michal Nánási worked more than 3 years at large tech company (over 1 billion of registered users) as software engineer, mostly working on protecting users from high-volume abuse with, and without machine learning.
Speech data mining: Not Yet Ready for Retirement (Honza Černocký)
Speech data mining is an important discipline of machine learning. In addition to classical ML and computer science components, it also sources other sciences such as physiology, lexicography, and phonetics, making it a funny inter-disciplinary domain. Similarly to other ML sub-fields, it has been turned upside-down in the recent years by the massive use of neural networks. Although their use in speech dates back to 2000, it is only around 2010 they took the ground and started to dominate the field.
Brno speech group has been through all these changes, sometimes following the others, sometimes defining the history. The talk will cover these developments as well as some new research trends.
Honza Černocký is Associate Professor (Docent) and Head of the Department of Computer Graphics and Multimedia at the Faculty of Information Technology, Brno University of Technology (FIT BUT). He also serves as the managing director of BUT Speech@FIT research group. He was with ESIEE Paris, France and OGI Portland, Oregon, USA. His research interests include signal processing and speech data mining (speech, speaker and language recognition, keyword spotting and spoken term detection). He was PI and co-PI of several local, European and US-funded projects. He served as co-chair of IEEE ICASSP 2011 in Prague and co-organized IEEE ASRU 2013 in Olomouc. As faculty member of FIT BUT, he teaches signal-processing, pattern recognition, and speech related courses. Honza co-founded Phonexia in 2006 and helped launching two other speech start-ups.
Time series prediction with Neural networks: Applications, Pitfalls and Beyond (Rudradeb Mitra)
One of the key application of AI and IoT is making time-series predictions. Companies are using predictive analytics in various ways – from predicting customer buying behavior to predicting health risk to predicting potential future accidents.
In this talk, the speaker will explain using three cases how are companies combining time series data and Neural networks to do predictions. Then he will explain what are the pitfalls in the approach of using Neural networks for predictions and what other kinds of algorithms are being or can be developed to overcome the hurdles.
Rudradeb Mitra worked in 4 European AI groups and had published 10 research papers on various AI topics including language processing, semantic web, 5th generation languages, and multi-agent planning. After finishing his Masters from Univ. of Cambridge he went on to build 4 startups – two in Silicon Valley, one in UK and one in Belgium. One of his startups was working on predicting potential future car accidents. These days he spends most of his time speaking and writing on topics related to Artificial Intelligence, startups and sales.
ML Engineer’s look at insurance (Peter Zvirinský)
Insurance is one of the oldest applications of mathematical modeling. The first statistical models designed specifically for insurance date back to the 18th century. Ever since has insurance modeling been often considered a separate topic from general mathematics/statistics and it is usually taught as a separate field of study on most universities.
This talk will present an ML Engineer’s look at insurance and you will find out what makes insurance so different from other typical ML applications. In the course of this talk all components of a typical insurance model will be broken down and you will be shown methods which can be applied for modeling each component. You will also find out what a tail risk is and why it is so important to insurance companies. Lastly, you will be shown why sometimes using a traditional statistical approach might still be prefered to a modern ML approach, and how one wrong assumption regarding your data can lead you to completely erroneous conclusions.
Peter Zvirinský is a Machine Learning Engineer at CEAI, where he works on TowerStreet, which is CEAI’s cyber insurance venture. He is also pursuing his PhD in Theoretical Computer Science on Faculty of Mathematics and Physics, Charles university. Prior to CEAI, Peter worked as an ML researcher at Seznam.cz’s online advertisement department.
Cesta vidiaca autá, ktoré po nej jazdia (Adam Herout)
„Lidi musí od rána do večera chodit do práce, aby si mohli koupit auto, které jim slouží k tomu, aby mohli jezdit do práce,“ řekl mi kdysi kamarád a možná na tom něco je. Doprava po silnici je tu s námi a je zdrojem radosti, úspěchu i mnohého trápení. Nemám recept, jak se toho trápení zbavit, vím ale, jak v tom trochu pomoct – zjišťovat o dopravě informace pomocí kamer. V přednášce promluvím o tom, v čem je potíž a jak s ní bojovat – pomocí strojového učení, geometrie a práce s daty.
Adam Herout – profesor na Fakulte informačných technológií VUT v Brne, kouč, „Někdy jen tak sedím a přemýšlím. A někdy jen tak sedím.“
Adam Herout je profesorom a vedúcim skupiny Graph@FIT na Fakulte informačných technológií VUT v Brne, kde se zaoberá počítačovým videním a rozšírenou realitou. Mimo akademickej činnosti je Adam Herout notorický zakladateľ a kibic startupov, ktoré sa točia okolo computer vision a machine learningu.
Okrem akadémie a biznisu Adam koučuje, a to hlavne profesionálov z oblasti IT.
3D Reconstruction in Wild (Martin Bujnak)
Photogrammetry is a seemingly magical process that can turn a set of photos into a detailed 3D model with only a PC. While not far from the truth, like most magic tricks, the details make the difference. In this talk Martin will cover how a pipeline for photogrammetry makes the difference between a poor reconstruction and an amazing one.
The talk will detail the technology behind and use of a specific photogrammetry tool, Reality Capture, and the use of that tool’s new SDK to extend the pipeline for better results. It will show how using an SDK to add instant, on-site, visual feedback during capture can make the process more reliable and successful.
While it might look magical, there are still many opened questions and opened problems to solve.
Martin Bujnak co-founded Capturing Reality which revolutionized process of 3D model creation from photos and laser scans.
He received his Ph.D. in computer vision from Center for Machine Perception in 2012 and the RNDr. and M.S. degrees in computer graphics and parallel computation from Comenius university in 2005. He pioneered and published several minimal solvers for camera pose estimation on major computer vision conferences and journals. He gained development skills in Caligary, Micrsoft live labs and Xbox.
Machine learning in adversarial environments for malware detection (Jakub Debski)
Recent discoveries in the field of neural networks and cheap parallel processing have enabled machine learning to be applied successfully to many data processing tasks. Its success has been so impressive that Machine Learning became a buzzword, and now every company tries to jump onto a “machine learning bandwagon” in order to be perceived as a “high-tech”. Unfortunately, too much marketing leads to too much gibberish.
This talk addresses recent claims from startup vendors and presents reasons why Machine Learning is not a silver bullet for malware detection. In IT security we have to contend with evasion techniques that are being developed in parallel to detection methods, and machine learning is no exception. The talk covers recent observations and results in this field and looks at evasion techniques from a security company’s perspective. It also explains why human oversight is still crucial in the classification process and why in adversarial environment a signature-less solutions face the same issues as the “old style anti-virus”.
Jakub Dębski, living currently in Bratislava, originally from Poland. Almost 20 years of professional experience in IT security out of which the last 10 years in ESET, currently as Head of Core Technology Development. Did master thesis in using neural networks for malware detection in Military University of Technologies / Warsaw. Since then actively tracking progress of machine learning and introduced the topic to ESET for classification of malware.
Seznam.cz – Vyhľadávanie obrázkov z (nielen) českého internetu (Lukáš Vrábel)
Lukáš Vrábel so Seznamu.cz nám priblíži, ako funguje Vyhľadávanie obrázkov z (nielen) českého internetu. Môžete sa tešiť na prelet históriou obrázkového vyhľadávania na Sezname – od jednoduchých textových prístupov až po hlboké neurónové siete.
Lukáš má v Sezname na starosti menší výskumný tím, ktorý pracuje na realizácii rozličných úloh strojového učenia od analýzy textu a webových stránok až po rozpoznávanie obrazu, takže sa máme na čo tešiť!
Využitie machine learningu v tvárovej detekcii a verifikácii (Marián Beszédeš)
Marián bude rozprávať o tom, kde všade pri spracovaní obrazu tvárí môžeme použiť rozličné techniky ML a aký bol ich vývoj. Zameria sa na techniky súvisiace s (hlbokými) konvolučnými sieťami aplikovanými na problémy tvárovej detekcie a verifikácie. Súčasne uvedie aj skúsenosti s ich optimalizáciou a reálnym nasadením (ladenie rýchlosti, vývojové postupy, použitý SW).
je computer vision team leader v Innovatricse
, kde sa zaoberajú problémami detekcie, trackingu, segmentácie a rozpoznávania tvárí. Touto problematikou sa za zaoberal už v rámci inžinierskeho a doktorandského štúdia na FEI STU. V minulosti sa zaoberal aj BIG data processingom v doprave a telekomunikáciách.
The role of optimization in Machine Learning (Dominik Csiba)
Optimization is a key component of most of the machine learning methods. Current problems often feature large amounts of high-dimensional data fitted by a complex non-convex model, such as deep learning, which are very difficult to learn efficiently. Novel approaches arise every month to tackle these problems, but the training time is often still unsatisfactory. In this talk, I will present a few scenarios, where optimization is the main bottleneck – featuring questions such as: “What kind of problems are difficult for deep learning?”. The talk is aimed for general technical audiences, everybody is welcome!
Automatické budovanie modelov pre časové rady s aplikáciou v energetike (Ján Dolinský)
Ján Dolinský zo spoločnosti tangent.works vysvetlí, prečo je automatické budovanie modelov tak atraktívne pre priemysel a prečo sa v tangent.works rozhodli pretaviť vedomosti a skúsenosti do produktu, ktorý sa nazýva TIM – Tangent Information Modeller. Ján bude rozprávať aj o informačnej geometrii, ktorá umožnila vybudovať tento produkt.
Extrakcia diskriminačných kľúčových slov (Márius Šajgalík)
Márius Šajgalík momentálne robí PhD v odbore Inteligentné informačné systémy na FIIT STU v Bratislave. Jadrom prezentácie bude reprezentácia textových dokumentov vo forme kľúčových slov pre účely kategorizácie a nové architektúry neurónových sietí, ktoré možno využiť nielen na extrakciu kľúčových slov bez učiteľa.
Budovanie NLP klasifikátorov (Vlado Boža)
Vlado Boža momentálne robí PhD v bioinformatike, účí ML na FMFI a okrem toho pôsobí ako machine learning lead v Black Swan Rational a aktuálne aj v CEAI Slovakia. Príde porozprávať o budovaní NLP klasifikátorov – od logistic regression po deep networks, od riešenia Kaggle súťaží po riešenei aktuálnych problémov v CEAI. Hovoriť bude aj o distant supervision a získavaní trénovaích dát a ukáže nám niekoľko efektívnych trikov na uľahčenie roboty.
Máš nápad na prednášku? Ozvi sa nám na firstname.lastname@example.org