Adrian Scoica

The Wanish Project

In October 2021, inspired partly by the contrast between how I experienced the city of Zürich before, during, and in the closing stages of the COVID-19 pandemic, I started a personal project for sanitizing videos by allowing the user to remove moving objects from it (including people and moving cars). The working name I gave to this project is Wanish (inspired by the popular household detergent product), and I tested the alpha launch on scenes from around Zürich.

I am attaching concrete examples of outputs from the project (taken at night in the first instance, order to be able to guarantee the stability of sky lighting conditions).

Between June 15th and June 26th 2015, I designed, organized, and taught a summer school on the topic of Machine Learning for Natural Language Processing (ML4NLP) at the "Politehnica" University of Bucharest. The workshop comprised two tages: (A) a taught component which took place between June 15th and June 19th with two hours of taught lectures daily, followed by a one-hour hands-on practical session in which the participants solved a small problem in order to apply the notions learned, and (B) a weeklong hackathon which took place between June 23rd and June 26th, and during which a selected team of participants worked under my supervision to create a prototype of a post-OCR text regeneration tool.

In total, 13 applicants were selected to participate, with background ranging from High School level to University Lecturer. The feedback was overwhelmingly positive, with an overall average score of 4.6 out of 5 across the feedback forms.

The main page for the summer school (which includes links to all the presentation slides from the first week), can be publicly accessed here.

My MPhil Research Project was focused on using unsupervised and weakly-supervised Machine Learning to Natural Language Processing, under the supervision of Dr. Diarmuid Ó Séaghdha and Prof. Stephen Clark.

In my thesis, titled "Multilingual Generative Models for Selectional Preference Learning" (click thumbnail to download PDF), I investigated the use of Latent Dirichlet Allocation (LDA) for inducing plausibility estimates specific to the selectional preferences of Verb-Subject and Verb-DirectObject pairs in English (and verified state-of-the-art performance on three other European languages), and I tested the feasibility of Vector Space Alignment to transfer the estimates from resource-rich European languages like English (for which dependency parsers can be trained), to languages which do not benefit from a large body of research: German, Spanish, and Romanian.

As part of my research, I produced clean, dependency-parsed corpora from the non-listy German, Spanish, and Romanian Wikipedia articles in CONLL format, which I am releasing below. Please read the "README.txt" file in each archive for details about how the text was extracted and processed, as well as for information regarding licensing.

For evaluation purposes, I also created the first test dataset for selectional preference estimation in the Romanian language, by eliciting responses online from native raters. You can access the datasets in either PNG or CSV format.

The methodology of compiling the dataset is detailed in the main body of the dissertation. If you wish to publish results based on this dataset, please contact me by email first.

The code developed as part of the project is hosted publicly on my BitBucket account.

In 2013, I carried out a study (click thumbnail to download PDF) on the problem of email spam classification in order to verify the empirical claim that the performance of word-based classifiers as a function of the leading K tokens of an email saturates quickly with increasing values of K. In doing so, I built on the work of (Çıltık and Güngör, 2008), and tested Multinomial Naive Bayes (MNB), Support Vector Machine (SVM), Bayesian Logistic Regression (BLR), and Interpolated Language Models (ILM) approaches on the GenSpam corpus.

The study concluded that the MNB, SVM, and BLR methods are robust to decreasing text message length down to approximately 100 tokens, while ILM is robust down to approximately 20 tokens.

Upon graduation, I wrote my Bachelor Thesis on the topic of automated plagiarism detection in specialized corpora (academic papers in the field of Computer Science), under the supervision of Dr. Traian Rebedea and Prof. Răzvan Rughiniș.

In my thesis, titled "The AuthentiCop System for Plagiarism Detection in Specialized Corpora: Algorithms and Data Processing" (click thumbnail to download PDF), I described building an automated pipeline which can perform plagiarism detection based on the Encoplot algorithm (Grozea and Popescu, 2011). Together with Filip Buruiană, we wrote a paper on the topic of the thesis which got the best paper award at the Student Scientific Paper Session of the "Politehnica" Unversity of Bucharest, 2012.

During my undergraduate years, I also acted as a Teaching Assistant for a number of courses at the Politehnica University of Bucharest: Operating Systems Usage, Computer Programming, Data Structures, and Algorithm Design. As a teaching assistant, I taught during laboratory classes, wrote laboratory exercises, came up with homework assignment and final exam questions, wrote tutorials, and managed course repositories. I am publicly releasing some of the work below.

The repository also contains the source code I wrote for the automated grader and visualizer for the second assignment of the 2011-2012 academic year, which required students to write an engine that can beat our AI at the Connect4 board game. Two video tutorials explaining how to use the visualizer to test hand-written AI bots for Connect4 are given below (Romanian language only).

As first a TA, and later the head of the TA team for the Computer Programming Course, among my typical duties I adopted and maintained an open source online judge platform for the course laboratories (which is live here). The source code for the platform belongs to the popular Romanian competitive algorithm design online judge Infoarena. By December 2014, the judge reached 638 registered students.

A short video tutorial instructing the students how to use the website is given above (Romanian language only).

In Fall 2011, I participated in the Google AI Challenge together with a team of 3 colleagues from my undergraduate course. We designed and implemented an AI bot that would control a swarm of ants as they forage for food and wage war against opponent swarms on the map. In the final championship, our bot ranked 63 out of a total of 7,897 teams from all over the world. You can watch an online demo of our AI bot at work below.

In 2010, I led a team to design a Chess engine in C++, which at the end of the year won the tournament organized as part of my Algorithm Design class. We developed the engine to be compatible with the XBoard Chess platform from the GNU Foundation. The implementation was based on an Alpha-Beta prunned NegaMax algorithm, with added support for Quiescence Seach, custom heuristics, and a database of openings from famous chess championships to buy us time in the early stages of the game. We implemented the algorithms over our custom bit-wise representation of the gameboard, hence naming the engine "BitBoard". You can watch an online demo of the engine below.

In 2011, upon becoming a Teaching Assistant for the Algorithm Design course at the Politehnica University of Bucharest, I promoted the use of C++ at the labs, and I wrote a Crash Course in C++ for Java Programmers handbook (click thumbnail to download PDF - Romanian language only) designed to integrate seamlessly with the OOP curriculum at the university. The hanbook proved successful with later generations as well, and was referenced externally, on websites such as www.itassistant.org.

Following its success, I was later invited in 2012 to teach a Crash Course in C++ workshop at THALES Systems Romania.

In 2010, a year after having graduated from the first edition of the ROSEdu Community Development Lab (CDL) myself, I taught a crash course in C++ Templates and the Standard Template Library (click thumbnail to download PDF slides - Romanian language only). The course was a success, which prompted me to return and teach it the following year as well.