FIS - Faculty of Information studies

Research seminar

Research seminar

Our Faculty holds regular research seminars where we seek to have as diverse speakers as possible, covering all research fields related to our work. Below is the list of recent seminars.

Seminar 13: Kognitivna robotika - novi razvojni izzivi

Date: 14. 6. 2016
Speaker
: prof. dr. Bojan Jerbič s Fakultete za strojništvo in ladjedelništvo Univerze v Zagrebu
Abstract (in Croatian):​ Ako želimo robote integrirane u naš svijet moramo biti u stanju razviti robote koji će razumjeti nas i našu okolinu. Iako su još Leibniz i poslije Kant, pa i drugi sljedbenici, zagovarali računarsku teoriju uma (tzv. ”computationalism”), svojevrsnu analogiju uma i mozga s računarskim programom i računalom, prema Gödelovom ”teoremu nepotpunosti” proizlazi da ljudski um nema uporište u Turingovu stroju ili teoriji automata. Ipak, novi pogledi na teoriju uma otvaraju ortogonalne pristupe oblikovanju kognitivnih modela temeljenih na novim teorijama prezentacije znanja i znanja o znanju, koji polaze od bioloških i socijalnih premisa. Visoka razina apstrakcije znanja, učenja  i zaključivanja uvjet su za realizaciju uistinu inteligentnih i autonomnih robota. Stvarno inteligentan robot treba biti sposoban izgraditi model sebe samoga. Nove primjene podrazumijevaju integraciju ljudi i strojeva, ne samo u pogledu suradnje, već i u pogledu biološke integracije. To ujedno znači da će se sve više  koristiti biološki sustavi za rješavanje tehničkih problema kojima nismo dorasli. U izlaganju će biti istaknuti glavni izazovi u razvoju kognitivnih strojeva, od semantičke percepcije, učenja, interakcije do tehničke koncepcije sustava. Posebno će biti predstavljena originalna istraživanja zasnovana na evolucijskim metodama, induktivnom učenju te vjerojatnosnim modelima oblikovanja robotskog ponašanja kroz interakciju robota i ljudi. Ostvarivanjem kognitivnih funkcija više razine stvoreni su uvjeti za primjenu robota u iznimno zahtjevnim zadaćama kao što je neurokirurška stereotaktička navigacija. Pored glavnih dostignuća projekta RONNA (”Robotic Neuronavigation”), predstavit će se bionika, kao jedan od glavnih znanstvenotehnoloških i etičkih izazova današnjice, adresirajući buduće smjerove istraživanja.

Seminar 12: (free discussion on nowadays popular topics in bioinformatics and computational biology)

Date: 1 marec 2016
Speaker: Kristina Ban, FIŠ. ​​​
Abstract: This seminar is intended not as a usual lecture, but rather as a free discussion. Kristina Ban will open and moderate the discussion on nowadays popular topics in bioinformatics and computational biology, mainly referring to biological networks (protein-protein interaction networks, metabolic networks, phylogenetic trees, gene expression networks), including different computational methods used in this vibrant field, such as network alignment. This will be the first part of a series ofseminars/discussions on this topic.

Seminar 11: Analiza družbenih omrežij: sedanjost in prihodnost

Date: 3 februar 2016
Speaker:
 
​​Robert Kopal, Visoka šola Algebra, Zagreb​, Hrvaška
Abstract:​ Social network is formally defined as a group of interconnected "individuals". These individuals are connected by one or more types of relations whose patterns occupy the attention of scientists and researchers. They relate to each other by either some kind of cooperation, or competition/conflict, and are referred to as entities or actors. Their relations can be depicted by a graph in which each entity (actor or individual) is depicted as what is in graph theory called a node, and their relations are depicted as links (ties). Social Network Analysis (SNA) is the structured analysis of social networks. Using mathematical tools and computer software it helps measure different types of variables related to the distance between the nodes (vertices) and the types of their associations, or links (edges), in order to determine the degree and type of influence one vertex has on another. SNA helps identify hidden associations and degrees of influence between the dots. The practical SNA application will be demonstrated in the following areas: telco industry, banking, national security, HR.

Seminar 10: Globoko vzpodbujeno (strojno) učenje v robotiki

Datum: 13 januar 2016
Predavatelja:
 Matjaž Omladič in Blaž Sobočan, Inštitut Jožef Stefan in Fakulteta za matematiko in fiziko Univerze v Ljubljani 
Abstract: The lecture will present the work in progress on robot control by joining efforts of existent group in the robotics department at IJS with the newly joined mathematicians. We concentrate on applying neural networks to the usual methods of reinforcement learning trying to get a better result. Most of applications of neural networks so far are limited to vision problems and various tries to use it in controlling robots did not turn our to be particularly successful. A new approach in neural networks, given also a new name - deep learning (possibly to forget about the less successful applications of the method), may deserve another try since a number of successful applications has been reported in various fields. This Summer we started by applying one of the deep learning methods developed as an optimization method. We developed a combination of these techniques with reinforcement learning by viewing it as an optimization problem. Later we found an even better method of deep learning for our purposes, namely the method of autoencoders. These were first developed for reproduction of images and aimed at reducing the number of starting nodes (visible units - possibly representing certain features) to a much smaller number of nodes (hidden features) but than going back to the starting number of nodes in order to reproduce the starting image with a smaller number of hidden features. This method was applied to a simulator and gives interesting results. We believe it deserves to be tested via a more realistic experiment and may lead to a publishable paper. The leading investigator is dr. Matjaž Omladič, supported by dr. Bojan Nemec and dr. Aleš Ude. Some work was done by a master student at Faculty of Mathematics, Physics and Mechanics, Blaž Sobočan, who will also present some of the tests using the Python libraries.

Seminar 9: Promocija slovenske znanosti skozi medijsko mrežo Metina lista

Date: 13 januar 2016
Speaker:
 Nataša Briški, so-ustanoviteljica in urednica medijske mreže Metina lista
Abstract: Seminar bo priložnost za predstavitev medijske mreže Metina lista (Metina lista, Meta Dekleta in Meta Znanost), medijskega start-upa, ki se v svojem delovanju osredotoča na aktivno državljanstvo, podatkovno novinarstvo in promocijo znanosti. Avtorji vsebin so strokovnjaki s svojih področij, ki odpirajo teme, pogosto prezrte v tradicionalnih medijih. Poseben poudarek bo na predstavitvi projekta Meta Znanost - portala ZA boljše razumevanje znanosti, ZA poljudnejšo razlago kompleksnih znanstvenih vprašanj in ZA popularizacijo znanosti. Podrobneje pa bo predstavljen tudi najnovejši projekt Metine liste, to je vizualizacija slovenskih znanstvenikov in znanstvenic, ki delujejo v tujini. S pomočjo različnih podatkovnih virov in CartoDB, orodja za geolociranje, geografsko prikazujemo slovenske znanstvenike po svetu, kje delujejo in kaj raziskujejo. V zaključku predstavitve več tudi o možnostih sodelovanja FIŠ z medijsko mrežo Metina lista.

Seminar 8: The construction of an annotated corpora in Slovenian language and the evaluation performance of sentiment based classification techniques

Date: 15 December 2015
Speaker:
 
Jože Bučar, Faculty of Information Studies
Abstract: The Web today is a growing universe of websites and a huge repository of structured and unstructured data. Enticing as it is with varied and freely accessible data, it is a remarkable source of data for every data scientist. Especially when dealing with textual data, language scientists are more frequently turning to the Web as a source of language data, precisely because it is so enormous, because it contains facts, emotions and opinions, which we can extract, or simply because it is free and instantly available. In this talk we introduce the methodology and tools that were required for their construction. Web crawlers retrieve the content of web pages in Hyper Text Marup Language (HTML) format from the news archives. The annotation process was carried out on three levels independently, i.e. document-level, paragraph-level, and sentence level. The annotated corpora contain more than 10,000 documents. Afterwards we provide some examples of its use, we evaluate performance of sentiment based classification techniques, and finally, we estimate the proportion of negative, neutral, and positive news in web media.

Seminar 7: Modeling dynamical systems with machine learning: The constant struggle against overfitting

Date: 8 December 2015
Speaker:
 
Ljupco Todorovski, Faculty of government, Univeristy of Ljubljana
Abstract: Modeling dynamical systems from empirical data is a genuine inverse problem: Given observations of the system dynamics, one looks for a model the most appropriate structure and parameters that explain these observations. The problem naturally fits the framework of supervised machine learning that aims at training a model from data. However, supervised methods are often prone overfitting -  in attempt to capture all the information from the data, the model becomes so detailed that it ends up having poor predictive performance on new data unseen in the training process. I will present our method of dynamical systems modeling, referred to as process-based modeling, emphasizing our approach to overfitting avoidance. In this context, I will also present empirical results evaluating different approaches to overfitting avoidance, ranging from classical minimal-description length principle that introduces general bias towards simpler models, through reducing the variance by using ensembles of models, to strengthening the bias by using domain-specific knowledge. The last approach relies on adding expected domain-specific properties of the model dynamics to the objective function that measures the fit of the model behavior to the training data. The ability of the method to take into account such properties of the model behavior have a very nice and somehow surprising consequence: the presented method can also address the new and practically relevant task of designing systems that produce behaviors with given desired properties. The presented work is a result of a collaborative research with Sašo Dzeroski, Nikola Simidjievski and Jovan Tanevski from Jožef Stefan Institute, Will Bridewell from Naval Research Laboratory and Pat Langley from University of Auckland.

Seminar 6: Zakaj rabimo hipotezo v sodobni znanosti?

Date: 26 novembar 2015
Speaker:
 
Dejan Jelovac, FIŠ
Abstract: Predavanje ima za svoj glavni namen, da razblinja zelo razširjeno zmotno in površno prepričanje o tem, da je resnico v znanosti treba preprosto iskati v »raziskovanju dejstev«: dovolimo naj facti bruti govorijo sami! To se morda lepo sliši, vendar ne drži, saj se nobeno raziskovanje nikoli ne more niti začeti preden se ne začuti neka težava oz. problem v praktični in/ali teoretski situaciji! Hipotezo znanstvenik potrebuje kot anticipacijo rešitve odkritega problema/težave. Zaradi tega bo predavanje poskusilo osvetliti najprej sam pojem in poslanstvo hipoteze, njeno vlogo in mesto v moderni znanosti. Temu bo sledila razlaga formalnih pogojev adekvatnosti hipotez kot glavna poanta predavanja. V zaključku se bomo posvetili razmisleku o prihodnosti uporabe hipotez v znanstvenem raziskovanju.

Seminar 5: Evolucijsko optimiranje

Date: 18 novembar 2015
Speaker:
 Gregor Papa, Inštitut Jožef Štefan, Ljubljana
Abstract: Evolucijsko optimiranje je stohastična hevristična metoda iskanja optimalnih vrednosti. Mnogo optimizacijskih problemov je tako zahtevnih, da bi za njihovo reševanje porabili natančni (deterministični) algoritmi preveč časa. Evolucijski pristop uporablja preiskovalne algoritme, ki posnemajo načela biološke evolucije. Pri tem gre za sočasno iskanje več rešitev istega problema in kombiniranje posameznih rešitev za izdelavo nove rešitve. Vključuje dve fazi iskanja: preiskovanje in izkoriščanje. Pri preiskovanju čim večjega dela prostora rešitev skuša identificirati obetavna področja, na katerih nato izvaja podrobnejše postopke izkoriščanja. Genetski algoritmi so neposredna optimizacijska metoda, ki temelji na mehanizmih evolucije in naravne genetike. Navdihnjeni so z naravno selekcijo, ki vodi k preživetju in nadaljnjemu razvoju najboljšega posameznika. Z uporabo verjetnostnih mehanizmov za zniževanje cene ima ta metoda veliko verjetnost, da najde globalno optimalno rešitev problemov z več spremenljivkami, ki imajo v splošnem več lokalno optimalnih rešitev.

Seminar 4: Nevtralna teorija molekularne evolucije in vloga redundantnih genov pri evolucijskem računanju

Date: 29 oktobar 2015
Speaker
Vida Vukašinovič, Inštitut Jožef Štefan, Ljubljana
Abstract: Algoritmi evolucijskega računanja, tj. algoritmi navdahnjeni z biološko evolucijo, zavzemajo pomembno mesto znotraj optimizacijskih metod, namenjenih reševaju težkih optimizacijskih problemov. Evolucijski algoritmi uporabljajo enake principe, kot jih srečamo pri evoluciji v naravi. To so reprodukcija, mutacija, rekombinacija (križanje) in selekcija. V naravi se med osebki razvije tekmovalnost za naravnimi viri, ki omogočajo preživetje. Ta tekmovalnost povzroči, da imajo pri preživetju in parjenju prednost bolje prilagojeni osebki. Naravni izbor poskrbi, da se bodo z večjo verjetnostjo razmnoževali posamezniki, ki so bolje prilagojeni okolju. Razmnožujejo se s postopkom križanja, pri katerem potomci podedujejo lastnosti svojih staršev. Pri križanju pa se lahko zgodijo tudi napake, kar imenujemo mutacije. Na ta način se razvijajo nove populacije, ki so raznolike in dobro prilagojene okolju. Takšen postopek evolucije posnemamo tudi pri evolucijskem računanju. Po Darwinovi teoriji evolucije naj bi torej vodilno vlogo prilagajanja osebkov na razmere okolja odigrala naravna selekcija. Japonski biolog Kimura pa v svoji nevtralni teoriji molekularne evolucije trdi, da se mutacije pojavljajo mnogo pogosteje, kot je veljalo prepričanje, in da so ravno številčne nevtralne mutacije tiste, ki povzročajo največ evolucijskih sprememb. Potencial nevtralnih omrežij pri uvajanju alternativnih poti v evolucijskem razvoju in posledična izboljšava kakovosti iskanja je glavna motivacija za uporabo redundantnih reprezentacij v evolucijskem računanju, ki so tema tokratnega seminarja.

Seminar 3: Statistical signal processing for condition monitoring of electrochemical energy systems – Application on PEM fuel cells

Date: 7 October 2015
Speaker: Pavle Boškoski, Jožef Stefan Institute, Ljubljana
Abstract: Despite the maturity of fuel cell development, the supporting systems that influence their optimal usage, reliability and longevity still require substantial development. Such supporting systems require a set of algorithms and hardware subsystems for control, condition monitoring (CM) and power conditioning. Achieving such a system four criteria should be met: (i) such a system should be capable of performing sufficiently fast and accurate estimation of the fuel cell CM, (ii) it should allow straightforward integration without any prior characterization of the fuel cell system, (iii)its output should be describe the overall health of the fuel cell preferably in the interval [0,1] and (iv) the algorithms should be computationally efficient allowing embedded implementation. Various faults alter the PEM fuel cell impedance characteristic. Therefore, electrochemical impedance spectroscopy is employed for the purpose of CM. Currently, the most popular technique for estimating PEM fuel cell impedance employs sinusoidal waveforms. Notwithstanding the precise impedance measurements, such an approach suffers from long measurement period. Significantly, faster impedance estimation can be achieved by employing pseudo-random binary sequence (PRBS) as a perturbation signal. The impedance characteristic is computed using continuous wavelet transform with Morlet mother wavelet. With such an approach, EIS is performed in the frequency band from 0.1 Hz to 500 Hz within 60 seconds. By using PRBS excitation signal the impedance components among different frequencies become dependent complex random variables. Determining their distributions allows optimal selection of the decision thresholds of the CM system. As a result, the CM thresholds are calculated employing solely data acquired from the system in reference state of health and the desired false alarm rate. The overall fuel cell condition is estimated by fusing the information contained in particular impedance component through their joint cumulative distribution function based on copula functions. The output of the joint CDF can be directly used as an overall scale-free condition indicator. The proposed algorithms allow computationally efficient implementation. As a result, using these algorithms a complete solution for condition monitoring system based on modular DC-DC converter, 90 channel fuel cell voltage monitor and an embedded diagnostic algorithm. The complete solution has been evaluated on a 8.5 kW fuel cell power system.

Seminar 2: The influence of knowledge on processes within loosely coupled systems in the field of public safety

Date: 17 June 2015
Speaker: 
Jernej Agrež, Faculty of information studies in Novo mesto
Abstract:
 Between 2010 and 2014, the communities located in the Lower Sava valley experienced four flood events. The flood events occurred due to continuous rain in central, northeastern and southeast parts of Slovenia. The least threatening of the flood events included increased water level of the Sava and the Krka rivers, which isolated only a few houses from the rest of the community. The most devastating event caused several roadblocks, flooding the entire areas of the communities located close to the two rivers.To be able to understand the community learning in the flood threatened areas and to get insight, how emerged community knowledge influences flood response processes, we gathered meteorological and hydrological data from the Slovenian Environment Agency and data on the severity of the flood events from the Administration for Civil Protection and Disaster Relief database. We also conducted several semi-structured interviews with the residents in flood-endangered areas and extracted relevant information out of national defense documents, regional and local emergency standard operating procedures.Through the data analysis, we addressed following research questions: whether is it possible to define and assess processes in loosely coupled systems; whether is it possible to map and assess community knowledge; how community knowledge influences processes in the loosely coupled system; and finally, how knowledge based process pattern recognition could be used for ensuring public safety.

Seminar 1: Graph recoloring

Date: 9 June 2015
Speaker: 
Marthe Bonamy, Univesity of Montpellier
Abstract
A proper k-coloring of a graph is an assignment of one color to each vertex such that no two adjacent vertices have the same color, and at most k different colors are involved on the whole graph. Given two proper k-colorings of a graph G, is there a way to recolor G from one coloring to the other while recoloring one vertex at a time and ensuring that G is always properly k-colored? In how many steps? We will present various conditions on the pair (G,k) for this to be possible in few steps. We consider in particular graphs with no long induced path and graphs with no long induced cycle. This is based on joint work with Nicolas Bousquet (McGill University).