Academic Lectures

There will be three sets of lectures by the following distinguished lecturers.

Prof. Pablo Duboue

Natural Language Generation, Traditional Approaches and Research Directions

Abstract

Human communication involves language understanding and production. The understanding part receives most of the attention in Natural Language Processing (NLP), but the subfield of Natural Language Generation (NLG) has been researching solutions to the challenges of language production for more than 50 years. A common misconception is that NLG is simply the inverse of understanding. However, the problems are related but different, as NLG concerns with the issue of choice: given certain information to communicate, it needs to choose among a wide variety of forms given communicative intentions and an understanding of the shared information with the hearer. In these lectures, we will start discussing traditional approaches, deeply tied to research systems and linguistic theories. We will then dive into new deep learning approaches that are rejuvenating the field.

• Lecture 1: We will cover traditional (knowledge-based and linguistic) approaches used in early on in the field. This study will help us understand the study questions of the field and put the progress made using machine learning methods in perspective. Topics include: input to the NLG system, the role of intention and user model. NLG systems architecture and subtasks (content selection, content planning, sentence planning, aggregation, referring expression generation, lexicalization, surface realization, linearization), linguistic theories employed (Meaning-to-Text, Functional Unification Grammars, Systemic Functional Linguistics), problems addressed (weather, stocks, sports, cooking). I will also discuss my experience with the MAGIC medical report generation system.
• Lecture 2: We will discuss statistical approaches to NLG, with a focus on evaluation methods. Topics include: generate-and-test, over-generating grammars, scoring using a language model, genetic programming for content selection rules, learning to order pre-modifiers, knowledge-text corpora, issues when training from poor quality text, BLEU metric, ROUGE metric. I will also discuss my experience with the PROGENIE biography generation system.
• Lecture 3: We will discuss the main architectures proposed for using artificial neural networks for natural language generation problems. We will start with a review of deep learning concepts. Topics include: language embeddings, soft-max layers, beam search, RNNs, LSTMs, attention models, pointer networks, encoding structured knowledge, the WIKIBIO corpus and current research directions. I will also discuss my current work with the keywords4bytecodes reverse engineering of Java bytecodes project.

Biography

Dr. Duboue has 20 years of experience on AI / NLP research and development. After obtaining a doctorate degree at Columbia University in NY, he moved to IBM Research to pursue work on Question Answering, where he was part of the team that beat the Jeopardy! champions in 2011. Since then he has focused on consulting for startups, accelerating the access of technology to its potential users through his Vancouver, Canada company Textualization Software Ltd.

Prof. Giuseppe De Giacomo

Reasoning About Actions: From Automata to LTLf/LDLf Synthesis and Planning

Abstract

In this course we will study AI planning for goals expressed over finite traces, instead of states. We will look at goals specified in two specific logics (i) LTLf, i.e., LTL interpreted over finite traces, which has the expressive power of FOL and star-free regular expressions over finite stings; and (ii) LDLf, i.e., Linear-time Dynamic Logic on finite traces, which has the expressive power of MSO and full regular expressions. We will review the main results and algorithmic techniques to handle planning in deterministic domains, and especially planning in nondeterministic domains, both under full and partial observability. We will also briefly consider stochastic domains. Moreover, we will draw connections with verification and reactive synthesis. The main catch is that working with these logics can be based on manipulation of regular automata on finite strings, plus standard forms of planning.

• Lecture 1: Automata (NFAs/DFAs) seen as trace recognizers to formally describe and verify models — (90min)
• Lecture 2: LTLf/LDLf specification and goals (and connection with NFAs/DFAs) — (90min)
• Lecture 3: Planning and Synthesis in nondeterministic domains (through games over DFAs) — (90min)

Biography


Giuseppe De Giacomo is full professor in Computer Science and Engineering at Universita di Roma “La Sapienza". His research activity has concerned theoretical, methodological and realization aspects in different areas of AI and CS, most prominently Knowledge Representation and Reasoning, Reasoning about Actions, Generalized Planning, Autonomous Agents, Service Composition and Orchestration, Process Modeling, Data Management and Integration. He is AAAI Fellow, ACM Fellow, and EurAI Fellow.

Prof. Gavin Brown

Abstract

Lecture 1: How to Get it: The Crazy Rollercoaster that is your PhD
===
Wow getting a PhD is tough isn't it? The PhD rollercoaster takes you up, and down, and up again. This session will hopefully provide a guide to this process, assuring you that everything you're seeing is normal. We will cover both the emotional and intellectual side of a PhD, and cover some tips on how to move a little faster, and more efficiently, through this process.

Lecture 2: Ensemble Learning
===
Ensemble Methods are one of the most widely used principles in all of Machine Learning, applied from simple linear models up to deep networks with millions of parameters. This session has two aims: (1) to make you aware of the main algorithms/principles in ensemble learning, and (2) to make you think hard about how, why and where these methods will be useful in modern data science.

Lecture 3: Feature Selection
===
Feature Selection is central to modern data science, from exploratory data analysis to predictive model-building, people identify subsets of their explanatory variables (or "features") in order to (1) reduce computational overhead, (2) reduce risk of overfitting, and (3) increase interpretability in their data. This lecture will cover the main issues in the field, and conclude with a research topic, that of "stability" in feature selection, or how reproducible and reliable your experimental results are.

Biography

Gavin Brown is Professor of Machine Learning at the University of Manchester, UK, where he is also Director of Research for the School of Computer Science. He has worked in Machine Learning for 20 years, in areas such as ensemble learning, neural networks, and information theoretic feature selection. He currently leads a team at Manchester working on basic research sponsored by EU/UK government, and industrially/socially relevant projects sponsored by AstraZeneca, ARM Research, and the local Police forces. He is a keen public communicator, and very much enjoyed a series of appearances on the children's BBC TV channel to explain robots and Artificial Intelligence.

Guest talks on spotlight topics

Vahur Puik

Historic Estonia on a modern “Streetview” ajapaik.ee

Arnis Paršovs

Estonian ID Card and Recent Security Issues

Abstract

The talk will give overview of Estonian ID card and the recent security issues found therein.

Biography

Arnis Paršovs is Computer Science PhD student at University of Tartu. He holds MSc degree in Cyber Security from Tallinn University of Technology and University of Tartu. His research interests include applied cryptography especially in the context of Estonian eID ecosystem.

Tõnu Esko

Genomics and Personalised Medicine

Abstract


Biography