Title & Description

Supervisor(s)

Multilingual Lexical Simplification 

Lexical simplification is the task of replacing complex words or expressions by simpler synonyms in a context-aware fashion. Lexical simplification is useful to make texts more accessible to different types of users such as people with cognitive impairment. In this project the candidate will investigate current methods in lexical simplification and implement techniques based on current continuous vector representations and neural network architectures. The project seeks to contribute to our current research on multilingual text simplification at TALN. The MsC candidate will have available both a dataset for experimentation and simplification software already available for several languages.

Horacio Saggion

Beyond Abstracts: Generating Summaries of Scientific Texts  

Scientists worldwide face the problem of scientific information overload since the pace at which scientific articles are published is increasing exponentially. In this scenario, the possibility to access to a brief and complete overview of the contents of an article is essential to cope with the great amount of papers to consider. Often abstracts, which are published together with scientific papers, are too short and lack essential information for a complete assessment of the value of the research presented. New approaches to the creation of text summaries, which identify the fundamental contents of documents, constitute a useful instrument to create rich, structured and focused synthesis of the contents of a publication, thus providing new ways to deal with scientific information overload. This master project aims at developing techniques to produce summaries of scientific documents based on an analysis of its content with advanced linguistics tools. The project will investigate techniques to train summarization systems based on available annotated data. The MsC candidate will have available a dataset for experimentation and text processing and summarization libraries to carry out the project.

Horacio Saggion

Criticize or Praise? Citation Characterization in Scientific Papers 

Scientific texts do not stand in isolation, they are connected to each other by means of citations that identify the background on which a given scientific work stands. Citations are particularly important in assessing research output, mainly by means of reference counts (e.g. h-index). Besides citation counting, in recent years, citation semantics, concerning the characterization of the purpose of a citation in a text, started to gain momentum. In order to fully take advantage of citations to assess a piece of work, it is particularly important to understand why a piece of work has been cited in a given context (give credit, identify methods and tools, provide background, criticize, etc.). The characterization of the purpose of citations can have a significant impact in many activities related to the fruition and assessment of scientific literature including scientific text summarization, scientific information retrieval, paper / author recommendation, etc. This master project aims at developing systems to automatically detect the semantics of a given citation in text. The work will be based on the use of supervised techniques to classify citations using a variety of information sources arising from the linguistic and semantic analysis of scientific documents. The MsC candidate will have available both a dataset for experimentation and text processing and summarization libraries to carry out the project.

Horacio Saggion

Extracting the Science from Research Articles 

This thesis will study and develop machine learning techniques (preferably Deep Learning techniques) to extract different types of information from research articles. The work will be based on the development of supervised techniques for the identification of the following types of information: problem, technique, results, advantages, disadvantages, among others. The student will have available data and software to carry out the work.

Horacio Saggion

How do you feel listening? 

This thesis will study and develop machine learning techniques (preferably Deep Learning techniques) to identify the sentiment and emotions of people listening to music in social networks. The study will be based on the analysis of social media data collected before, during, and after concert performances. Contextual information will be used to improve the performance of current systems. The student will have available data and software to carry out the work.

Horacio Saggion

Summarizing Multimodal Content: the case of text and images 

This thesis will investigate the contribution of textual and non-textual information for the summarization of long articles which include multimodal information. Recent works have shown that classification systems can work better when information from multiple modalities is used. This has been little investigated for summarization.

Horacio Saggion

Social media monitoring for disaster risk reduction 

Our group has contributed to the development of social media monitoring tools for flood risk reduction within a platform developed by the Joint Research Center (JRC) and to be incorporated into EFAS. We would like to extend this tool to consider other types of risk including forest fires, earthquakes, storms, and so on. This topic is fairly practical and application-oriented, and requires solid knowledge of Python. Technologies to be mastered during the process include online/streaming algorithms for text clustering, deep learning for text classification, and the integration of these tools. A good starting point in terms of a base of code for the platform and training sets exist.

Carlos Castillo

Gender bias in online platforms 

This project will analyze and study potential gender biases of online platforms. These may include job search engines, e-commerce websites, search engines, and platform economy sites for which some data could be scrapped/obtained. The specific platform is to be decided with the master student. The work will involve data scrapping/collection, measurement through fair ranking/recommendation algorithms, and analysis of potential biases found. It requires solid knowledge of Python and desire to work within a framework of data feminism, which seeks to eliminate gender biases in data collection and data processing pipelines.

Carlos Castillo

Risk-Prediction Instruments in Criminal Justice 

The Justice Department of Catalonia has been using for several years structured risk prediction instruments, similar to COMPAS in the US or OASIS in the UK, to predict the risk of recidivism and violent recidivism. Correctional officers, psychologists, and social workers apply a questionnaire to juvenile offenders and to people who have been convicted for a crime (these are different questionnaires); the questionnaire is then processed using a model trained on thousands of previous cases, and an output is produced, which is then interpreted by the person applying the instrument. Detailed anonymized data is available for studying to what extent different instruments can be biased against different sub-populations (e.g., immigrants or children of immigrant). Additionally, we would like to explore the potential for new risk prediction instruments. The work would be done in collaboration with researchers on criminology at the University of Barcelona and at the Justice Department of Catalonia.

Carlos Castillo

Discrimination in house rental market 

This project will analyze and study potential discrimination occuring in the house rental market in an online platform. With a recent sociological study demonstrating the existence of certain levels of discrimination from the side of the real state agents, the goal of this project will be to understand how much this discrimination is reflected, mitigated or amplified by the recommenders systems that operate in online platforms. To develop the project, the student will study the existence of such inequities based on the ethnicity, gender and sexual orientation (between other features) of the applicants and how the recommendations of the platform vary depending on such characteristics. It requires solid knowledge of Python for data scraping, data collection, data cleaning for its analysis.

Carlos Castillo

David Solans

Multilingual Neural Natural Language Text Generation 

Deep learning models have become common for natural language generation. However, so far, they focus predominantly on the generation of isolated sentences from shallow linguistic (syntactic) structures. In the current thesis, a deep model for the generation of paragraph-long texts from generalized syntactic graphs will be explored. The model will, in particular, be able to capture co-references (such as, e.g., “Barcelona is a lovely city. I live here already for more than 15 years”) and the coherent structure of the narrative. The developed model will be tested on available large datasets for a number of languages.

Leo Wanner

Deep reinforcement learning-driven dialogue management 

In order to ensure a flexible coherent conversation between a human and the machine that goes beyond predefined information exchange patterns, advanced dialogue management strategies must be developed, which take the history and the goals of the conversation into account and which are able to handle interruptions, side sequences, grounding and other phenomena of a natural dialogue. Neural network-based Reinforcement Learning models have shown to have the potential to cope with these challenges. In the current thesis, such a model will be explored on a large dataset of movie dialogues.

Leo Wanner

Identification of the communicative intent of the speaker in spontaneous dialogues 

In order to adequately react to a statement of the user, a conversational agent must not only understand the content of this statement, but also identify correctly the intention of the user when they utter it. In the context of this thesis, a generic deep learning-based communicative intent identification (and classification) model will be developed. As training and development corpus, one of the available large-scale dialog corpora will be used. The model will be tested on out-of-domain dialogues.

Leo Wanner

Bootstrapping a multilingual collocation dictionary 

Collocations are idiosyncratic (i.e., language-specific) expressions such as “take a walk” and “ask a question” (cf. in Spanish dar un paseo, lit. 'give a walk' and hacer una pregunta, lit. 'make a question' respectively) are a great challenge in both natural language processing and in second language learning - partially also because their meaning is not composed of the meaning of the isolated words that participate in the combination (thus, you don't 'take' or 'give' anything when you go for a walk). The goal of this thesis is to develop a deep learning (neural network) - based algorithm for bootstrapping a multilingual English - L2 dictionary of collocations from large corpora, assigning the meaning to the extracted collocations in accordance with a given typology. The work of the thesis will build upon existing example-based deep learning and word embedding implementations and multilingual corpora.

Leo Wanner

Luis Espinosa Anke

Second Language Teaching Assistant 

Spellcheckers and grammar checkers are already common. However, lexis checkers that control the correct use of words, collocations or idioms in our writings are still rather rudimentary. In the context of this thesis, an assistant for correcting the writings of second language learners with respect to their word choice will be developed. The work will start from an already available prototypical collocation checker for Spanish. The second languages from which one should be addressed are: Catalan, German, Greek, or Spanish. The work will be embedded in a large scale European project on migrant reception and integration.

Leo Wanner

Automatic novel analysis 

Automatic profiling of authors of novels with respect to their (first of all syntactic) style and/or preferred word choice or even identification of authors of novels has achieved a reasonable quality. This thesis will aim to complement the existing research. It will develop a model for the analysis of novel plots, including, e.g., the main characters and their profiles, the relations and types of interactions between them, the global time line, etc. The model will be trained and tested on the material from the Gutenberg corpus.

Leo Wanner

Juan Soler Company

Audio-visual singing voice separation 

Music source separation is the automatic estimation of the individual isolated sources that make up the audio mixture. Some of the applications are for example: automatic creation of karaoke, music transcription, or music unmixing and remixing. In this project, the goal is to separate vocals from instrumental accompaniment on an audio mixture by leveraging visual and motion information from the singer’s mouth. Deep neural networks that extract features from the video sequence will be explored and used in conjunction with an audio network in order to improve the audio source separation task by incorporating visual cues.

Gloria Haro

Process mining to understanding how teachers design learning activities 

Authoring tools and community platforms devoted to teachers (e.g. the Integrated Learning Design Environment) collect data about teachers’ action in the process of designing learning activities. In this context, the application of process mining techniques would bring light about how teachers design, i.e. which process they follow to gather inspiration by exploring designs created by others to which steps they follow in the authoring process. The TIDE group has developed several authoring tools and the ILDE community platform, which have been used by several teacher communities (two schools, teachers participating in professional development programs, etc). This project will consist in applying process mining techniques to these datasets, extract knowledge about how teachers design in each community and compare them.

Davinia Hernández-Leo

Ishari Amarasinghe

Intelligent Interactive Systems in Education 

Several topics related to the application of interactive and artificial intelligence techniques to the design of systems to support teaching and learning (includes adaptive and personalized learning, classroom orchestration, etc.).

http://www.upf.edu/web/tide

Davinia Hernández-Leo

Data Analysis in Education Technologies 

Several topics related to the application of data analytics techniques to the design of systems that support teaching and learning (includes community, teaching and learning analytics, interactive dashboards, etc.).

http://www.upf.edu/web/tide

Davinia Hernández-Leo

Conversational agents/chatbots for CSCL applications 

Conversational agents have been deployed into a variety of learning technology applications to enrich interaction between humans and machines. Tutorial Dialog Systems that employ Conversational Agents (CAs) to deliver instructional content to learners in one-on-one tutoring settings have been shown to be effective. This project focuses on extending this technology to collaborative learning settings. The student will focus on the development, integration and evaluation of conversational agents (following iterative design process) into a computer-supported collaborative learning (CSCL) tool called 'PyramidApp' which facilitates easy deployment of collaborative learning activities in classroom and distance learning settings.

http://www.upf.edu/web/tide

Davinia Hernández-Leo

Ishari Amarasinghe

Learning analytics for learning redesign and orchestration 

TThere are two problems that teachers face in their daily routines that would require data-driven educational technology solutions. First, the design and redesign of increasingly effective learning situations are currently not informed by indicators of the impact in learning of previous design realisations. Second, the orchestration of learning situations is a daunting task for teachers, that involves the monitoring, awareness and regulation of learning activities. Both problems stem from the fact that obtaining the adequate information and support required to make decisions about the (re)design and orchestration is out of reach for teachers. Learning Analytics (LA), which borrows methods from Artificial Intelligence (AI), can be considered a suitable approach to tackle both problems. However, it is unclear if the same LA solutions would be able to satisfactorily support both learning (re)design and orchestration. This TFM will analyse and compare the nature of both tasks and existing solutions supporting them, focusing on a case study that uses the same collaborative learning tool with LA to support both tasks.

Davinia Hernández-Leo

Ishari Amarasinghe

Understanding participant behaviours in Citizen Science online learning activities 

Citizen Science (CS) involve the collection and analysis of data relevant to solve research questions by members of the general public, usually as part of a collaborative project with professional scientists. In this master thesis project, the student will select a CS activity that is supported by technology (e.g., a CS devoted platform) to analyze how participants behave and interact with technology to collaborate in the endeavour. The analysis can target organizational/operational characteristics, scientific outcomes, individual/group learning, other success or failure indicators, etc., and societal aspects, related to the impact of those activities on society, such as gender, age, geographical and socio-economic differences; etc.

Davinia Hernández-Leo

Patricia Santos

Twitter as an educational community: identification and classification of educational tweets 

Several studies have shown that teachers use Twitter for professional development (PD) purposes and to develop a sense of professional community and reduce perceptions of isolation. Twitter has the potential to support the improvement of practice through grassroots continuing professional development. Recently, with the mandatory adoption of remote teaching due to the COVID-19 crisis, educators have found on twitter a good place to share educational resources, ideas for learning activities, messages with technical advice on the use of educational tools, etc. All this has generated a large amount of materials scattered throughout the network. This dispersion makes it difficult to get the maximum potential from all the knowledge generated and shared. For an educator, finding specific solutions for a specific need on Twitter is a challenge, as often the hashtags used are not consistent or specific enough as no standard categorization is used. This thesis aims to develop strategies for the automatic recognition and classification of educational information on Twitter. The strategies will ideally implement an incremental learning mechanism, which will allow for a continuous improvement of their performance.

Laia Albó

Pablo Aragón

Data-driven Forecasting of Atmospheric Pollution 

Reliable long-term forecasting of atmospheric pollution critically depends on meteorological conditions. Despite the high availability of meteorological and atmospheric pollution data, it remains unclear whether integrating meteorological predictions at the macro-level with data-driven models built from pollution measurements at the micro-level can lead to more accurate predictions, especially for predicting concentrations of nitrogen dioxide. This project aims to answer this question by learning state-of-the-art generative models developed at Fujitsu from data measured at the XVPCA, the Atmospheric Pollution Monitoring and Forecasting Network in Catalunya, in combination with macro-scale meteorological predictions. The research will involve learning such models, evaluating their accuracy in a real-world setting, and analyzing their potential for other tasks, such as simulating alternative atmospheric scenarios for subsequent analysis. This project is part of a collaboration between the Departament de Territori i Sostenibilitat de la Generalitat de Catalunya, the Centre de Telecomunicacions i Tecnologies (CTTI), and Fujistu.

The Atmospheric Pollution Monitoring and Forecasting Network (in CAT)

Vicenç Gómez

Where should I park Today? 

Reliable prediction of parking occupancy could result in considerable improvements in urban mobility. This project considers a rich longitudinal dataset of parking occupancy from several parking stations. Currently, this data is mainly used for monitoring purposes and it is typically discarded afterwards. The project will analyze the feasibility of using such data to build a predictive model for parking occupancy. This research will involve a preliminary statistical analysis to understand the raw sensor data in combination with parking validation data, building a predictive model from this data together with other exogenous factors such as weather, day of the week, etc., and evaluating the performance of such a model at different time-scales. This project will be developed jointly with the Autoritat del Transport Metropolità (ATM) de Barcelona.

Vicenç Gómez

Statistical Modeling of Online Discussions 

Online discussion is a core feature of numerous social media platforms and has attracted increasing attention for different and relevant reasons, e.g., the resolution of problems in collaborative editing, question answering and e-learning platforms, the response of online communities to news events, online political and civic participation, etc. This project aims to address, from a probabilistic modeling perspective, some existing challenges, both computational and social, that appear in platforms that enable online discussion. For example, how to deal with scalability issues, how to evaluate and improve the quality of online discussions, or how to generally improve social interaction such platforms.

Generative models of online discussion threads: state of the art and research challenges

Vicenç Gómez

Andreas Kaltenbrunner

Distributed control for teams of autonomous UAVs 

The aim of this project is to derive controllers for teams of unmanned aerial vehicles (UAVs). The focus will be on extending the current centralized algorithm to a distributed setting. The simulator used for the centralized version is implemented in ROS. Required MIIS courses: machine learning, autonomous systems, mobile robotics.

Real-Time Stochastic Optimal Control for Multi-agent Quadrotor Systems

Vicenç Gómez

Modelling the cooperative behaviours in real-time environments using Reinforcement Learning 

In this project we will model real-time experimental Game Theoretic tasks involving several agents using Reinforcement Learning techniques. The model will be based on Markov Decision Processes (MDP). The aim is to be able to make predictions on modifications of the experiments and to add increasingly complex features to the model including the prediction of other agent behavior and agent identity. Cooperative emerging behaviors will be studied, like for example in the presence of limited resources as in the Tragedy of the Common case where agents need to learn to consume resources in a controlled and coordinated way. Requirements: machine learning, autonomous systems

Vicenç Gómez

Martí Sanchez-Fibla

Generation and Detection of Deepfakes 

In the recent years many methods for face swapping and manipulation commonly known as "deepfakes" are rapidly evolving. The deepfakes make use of state-of-the-art techniques from deep learning and computer vision fields, making them increasingly harder to detect even for human evaluators. While there are legit applications of these techniques for creating deepfake videos in the audiovisual industry, they have the potential to abuse and to individuals, or propagate fake news. The focus of this thesis is first review the state-of-the-art and develop techniques for creating deepfakes and second, develop new algorithms to detect deepfakes and manipulated media, participating in the "Deepfake Detection Challenge", if results are satisfactory. The project will be carried out in Telefónica Research.

Vicenç Gómez

Ferran Diego

Carlos Segura

Program Synthesis for AI Planning 

Program synthesis is the problem of automatically generating programs, either from a declarative description or from examples such as input-output pairs. This project considers the problem of synthesizing programs for AI planning, in which the basic instructions are planning actions. Programs can also include instructions for control flow (i.e. conditional statements or loops) and for calling other programs. The purpose of the project is to extend an existing framework for program synthesis by implementing several alternative solution strategies, such as parallel execution and program compression, with the goal of improving scalability. A possible application of program synthesis is computer security, automatically deriving control rules that allow a system to detect and prevent external attacks.

Anders Jonsson

Sergio Jiménez

Novel Algorithms for Centralized Multiagent Planning 

Centralized multiagent planning is the problem of computing a joint plan for a team of agents that collaborate to achieve a goal. This problem has many applications in the real world, e.g. ride sharing, logistic planning and manufacturing. The aim of this project is to develop novel algorithms for multiagent planning based on an existing approach that reduces centralized multiagent planning problems to exponentially smaller approximate AI planning problems that can be efficiently solved using state-of-the-art planning methods. The novel algorithms will be tested in challenging scenarios from the Multi-Agent Programming Contest.

Anders Jonsson

Factored Markov Decision Processes 

Reinforcement learning (RL) has achieved several groundbreaking results in the last few years. Most state-of-the-art RL algorithms treat the state as a black-box, even though the state is usually made up of multiple components. There exists a theory for factored Markov decision processes (FMDPs) in which states and actions are represented as collections of factors. Crucially, some of these factors are conditionally independent of one another, which makes it possible to represent the MDP dynamics more compactly. The aim of this project is to study special cases of FMDPs that can be decomposed into smaller problems that can be solved individually. These solutions can then be combined to produce an overall solution to the original problem. If successful, this approach has the potential to reduce the learning effort by several orders of magnitude.

Anders Jonsson

Sadegh Talebi

Constrained optimization methods for computational optimal transport 

The framework of optimal transport addresses the problem of measuring distances between probability distributions. A rough definition of optimal transport distances can be given as follows: Given two probability distributions P and Q over the sets X and Y, and a cost function measuring transportation cost between elements of the two sets, the optimal transport distance between P and Q the total cost of transporting all mass from P to Q. This optimization problem can be formulated as a linear program called the Monge--Kantorovich optimal transport problem. This project follows up on recent progress made on computationally efficient algorithms for solving this LP, and particularly investigates the possibility of employing techniques for constrained optimization and saddle-point optimization for improving existing solutions. The project requires very strong mathematical skills, particularly in multivariate calculus and linear algebra. Knowledge of convex analysis and optimization is a plus, but not absolutely necessary at the current stage.

Sinkhorn Distances: Lightspeed Computation of Optimal Transportation Distances

Near-linear time approximation algorithms for optimal transport via Sinkhorn iteration

Gergely Neu

Thompson sampling for sequential prediction 

Thompson sampling is one of the most well-studied algorithms for a class of sequential decision-making problems known as stochastic multi-armed bandit problems. In a stochastic multi-armed bandit problem, a learner selects actions in a sequential fashion, and receives a sequence of rewards corresponding to the chosen actions. A crucial assumption made in this problem is that the rewards associated with each action are random variables drawn independently from a fixed (but unknown) distribution. The goal of this project is to do away with this assumption and study Thompson sampling in non-stationary environments where the rewards may be generated by an arbitrary external process. Precisely, the project considers the framework of sequential prediction with expert advice, and aims to show theoretical performance guarantees for this algorithm and/or analyze its perormance empirically. The project requires very strong mathematical skills, particularly in probability theory and multivariate calculus. Knowledge of convex analysis is a plus, but not absolutely necessary at the current stage.

Learning to Optimize Via Posterior Sampling

An Information-Theoretic Analysis of Thompson Sampling

Online Linear Optimization via Smoothing

Thompson Sampling for Adversarial Bit Prediction

Gergely Neu

Better algorithms for online linear-quadratic control 

Linear-quadratic control is one of the most well-studied problem settings in optimal control theory: it considers control systems where states follow a linear dynamics, and the incurred costs are quadratic in the states and control inputs. In recent years, the problem of online learning in linear-quadratic control problems have received significant attention within the machine-learning community. One particularly interesting development is the formulation of the control problem as a semidefinite program (SDP), which allows the application of tools from online convex optimization. The present project aims to develop new algorithms for online linear quadratic control based on this framework by exploring the possibility of using regularization functions that make better use of the SDP geometry than existing methods based on online gradient descent. The project requires very strong mathematical skills, particularly in multivariate calculus and linear algebra. Knowledge of convex analysis and optimization is a plus, but not absolutely necessary at the current stage.

Online Linear Quadratic Control

Online PCA with Optimal Regret

Gergely Neu

Implicit regularization methods for high-dimensional optimization 

This project aims at studying the regularization properties of various incremental optimization methods such as gradient descent, averaged gradient descent, and exponentiated gradient descent. While recent work has successfully uncovered relations between averaging schemes for gradient descent and L2 regularization, these results remain specific for the classical problem of linear least-squares regression. One branch of this project is concerned with generalizing these results to more general convex optimization problems. Another direction the project aims to explore is the regularization effects of other gradient-descent variants, and particularly the sparsity-inducing properties of exponentiated gradient descent. The project requires very strong mathematical skills, particularly in multivariate calculus and linear algebra. Knowledge of convex analysis and optimization is a plus, but not absolutely necessary at the current stage.

Exponentiated Gradient versus Gradient Descent for Linear Predictors

Iterate averaging as regularization for stochastic gradient descent

A Continuous-Time View of Early Stopping for Least Squares

Connecting Optimization and Regularization Paths

Implicit Regularization for Optimal Sparse Recovery

Gergely Neu

Error averaging in approximate dynamic programming 

Error-propagation analysis of approximate dynamic programming methods is a classic research topic with several implications on reinforcement learning algorithms. The most well-known results consider basic DP algorithms like value iteration and policy iteration, with several new results added in recent years regarding regularized counterparts of these methods. One discovery on this front was the discovery that certain regularization choices result in an "error-averaging" property that ensures a benign propagation of policy evaluation errors: roughly speaking, instead of accumulating the absolute errors linearly, these schemes allow the cancellation of positive and negative errors. This project sets out to understand this phenomenon in more detail by decoupling the averaging effect from the regularization effects as much as possible, thus potentially enabling the development and analysis of more effective error-averaging RL algorithms. The project requires very strong mathematical skills, particularly in linear algebra and probability theory. Knowledge of convex analysis and optimization is a plus, but not absolutely necessary at the current stage.

Approximate Modified Policy Iteration and its Application to the Game of Tetris

Leverage the Average: an Analysis of KL Regularization in RL

Munchausen Reinforcement Learning

Gergely Neu

Generalized information-theoretic generalization theory 

EThe generalization capability of machine-learning algorithms has been a central topic of ML theory since the inception of the field. Indeed, some of the most fundamental results of statistical learning theory concern the generalization error and give upper bounds on it in terms of various properties of the function classes used by the learning algorithm (for instance, its Vapnik--Chervonenkis dimension or its Rademacher complexity). However, these results are no longer applicable for modern machine-learning systems (such as deep neural networks) that use extremely large function classes of effectively unbounded complexity. To address these concerns, several alternative methods were proposed in the past years to characterize the generalization error in terms of quantities that do not only depend on the function class, but also properties of the training data and the learning algorithm. One particularly interesting line of work aims to bound these errors in terms of the mutual information between the dataset and the output of the learning algorithm, leading to tight bounds when the output of the learning algorithm is stable with respect to its input in an "information-theoretic" sense. This project aims to go beyond traditional notions of information-theoretic stability and explores divergence measures beyond the ones suggested by classical information theory. The starting point is a recent generalization of existing results from the perspective of convex analysis, with the eventual goal of identifying the right kind of divergence measures that lead to the tightest possible bounds for each combination of data distribution and learning algorithms. The project requires very strong mathematical skills, particularly in convex analysis, probability theory and information theory.

Information-theoretic analysis of generalization capability of learning algorithms

Generalization Error Bounds for Noisy, Iterative Algorithms

Reasoning About Generalization via Conditional Mutual Information

Gergely Neu

Learning and Planning in Simple Video-Games 

One of the big breakthroughs in AI during the last few years was the DQN algorithm that learned to play Atari video games directly from the screen using a combination of deep learning and reinforcement learning techniques. Neither DQN nor the systems the follow it, however, learn to play these games as humans do which is by understanding the video games in terms of objects and relations, and planning accordingly. The goal of the project is to make progress in that direction by 1) defining suitable high-level planning languages for modeling some these games, 2) defining planning algorithms for deriving the actions to be done in such games when the model is known, and 3) learning parts of the model from observed traces when the model is not fully known. In principle, we will work with symbolic models, complete or incomplete, and not directly with the information available in the screen. That would be a follow up step which goes beyond the work that can be done in a Master project. Some games to consider: Point-and-shoot games, Pong, Space Invaders, Pacman, etc. None lends itself to be modeled and solved by current planning languages and algorithms.

GVG-AI competition

VGDL: Video Game Description Language (used in GVG-AI)

OpenAI Gym

Planning with pixels in (almost) real time

Planning with simulators

Hector Geffner

Prediction of trabecular micro-fractures 

Bone fracture is a very local event. The fractured tissue has peculiar morphometric characteristics, however the prediction of this event is still an open question that lead every year to thousands of unpredicted or miss-treated patients. An ongoing study allowed us to study the morphometrical characteristics of trabecular fracture using image processing tools and micro-CT images. During this project we want to develop a classifier able to identify the weak region within the trabecular framework and predict its possibility to fail. The classifier will integrate information coming from the geometry and the mechanical behaviour of the trabecular structure.

Simone Tassani

Ranking Attributes for Information Gain in Forests of Decision Trees (data science / machine learning related project) 

Decision trees are one of the most well-known techniques in machine learning, data science, analytics and data mining. They are transparent: can be presented as rules for human understanding. With the availability of big-data, algorithms for construction of decision trees in a platform like MapReduce are a new challenge. This project consists of developing the parallelization algorithms and their implementation, perhaps for Hadoop when we consider forest of decision trees. This implementation will be focused on the applications of the forest of decision trees on privacy-preserving data mining on on-line social networks. We can provide references to recent literature on how the forest of decision trees are used to provide privacy alerts to users of OLSN. Another application is feature selection.

Vladimir Estivill-Castro

Manipulation with Pepper arm and fingers 

The project consists of developing the infrastructure and integrating motion planning and task planning algorithms for a Pepper robot to pick up boxes, like small cereal boxes. Currently, there is significant interest in the robotics community to combine these two types of planning to carry out tasks like cleaning a table with different objects. This kind of task is now common in the [email protected] challenges. The project should include elements of feedback between motion planning and vision. The idea is that vision may not recognise objects but suggest new positions which would enable the recognition of objects. Simultaneously, motion planning may be refined by feedback from vision.

Vladimir Estivill-Castro

Combining reasoning with robotic localisation 

There are many algorithms for robotic localisation in a field. However, these algorithms are really integrated with qualitative reasoning approaches. The most famous example of qualitative reasoning is Allen's interval algebra (and associated algorithms). The challenge is to investigate a spatial qualitative reasoning system over the robotic localisation in a soccer field to obtain strategic or tactical decision making that shows improved performance is the soccer field for RoboCup soccer.

Vladimir Estivill-Castro

Robot localisation in a soccer field 

There are many algorithms for robotic localisation in a space. However, these algorithms are really integrated with qualitative reasoning approaches. The most famous example of qualitative reasoning is Allen's interval algebra (and associated algorithms). The challenge is to investigate a spatial qualitative reasoning system over the robotic localisation in a soccer field to obtain strategic or tactical decision making that shows improved performance is the soccer field for RoboCup soccer.

Vladimir Estivill-Castro

Game playing Pepper 

The aim is to develop a demonstration of human-robot interaction by which a Pepper robot plays as naturally as possible a simple game, like tic-tac-toe on a fixed space, perhaps with rope and special pieces. The robot applies some localisation and some game strategy. The entire software is to be developed using model-driven development and finite-state machines as much as possible for coordinating the control. The robot shall be flexible in its speech and use localisation within the game space but not necessarily within the room. Sensor fusion to detect humans have completed their move is the main research challenge.

Vladimir Estivill-Castro

Ethical Pepper 

The aim is to develop a demonstration of human-robot interaction by which a Pepper robot learns ethical dilemmas and reason over ethical challenges. The robot shall apply several techniques including game theory or dialogue with humans to argue and explain ethical decisions. The goal is to investigate practical debates between a robot and humans regarding the alignment problem.

Vladimir Estivill-Castro