We organise monthly seminars in areas related to Data Science. To receive notifications about future events, please subscribe to our .
Title: Understanding Large Language Models and the Path to AGI
Speaker: Prof. John Abela
Date & Time: Wednesday, 5 November 2025, 12:00 (noon)
Location: Faculty of ICT, ICT Communications Lab (Level 0, Block B, Room 1)
Abstract
Neural Networks are often perceived by the general population as a form of magic, but at their core, they are essentially a structured sequence of mathematical transformations mapping an input tensor space to an output tensor space. Large Language Models (LLMs), such as ChatGPT, operate through a series of tensor algebra operations, leveraging vast amounts of data and computation. The true "magic" emerges not from individual calculations but from scaling—when models grow larger, they exhibit emergent properties that were not explicitly programmed. This talk explores the implications of scale in AI, drawing lessons from nature. Evolution did not grant humans 86 billion neurons and 100 trillion synaptic connections by accident; nature is economical, and the complexity of human intelligence is deeply tied to its capacity. The human brain's encephalization quotient—the ratio of brain mass to body size—exceeds that of any other primate, highlighting the importance of scale in biological intelligence.
A central question arises: Are human intelligence and consciousness Turing-computable? If intelligence is simply the product of sufficient capacity and complexity, then in principle, AI models, when scaled, should be able to achieve human-level Artificial General Intelligence (AGI). But does the nature of intelligence go beyond computation? What is the Kolmogorov complexity of human intelligence? The Chinese Room argument, proposed by philosopher John Searle, challenges the idea that syntactic manipulation alone is sufficient for true understanding. Meanwhile, philosopher and cognitive scientist Daniel Dennett’s theories on consciousness suggest that intelligence is just an emergent property of information processing, much like what we observe in modern AI models.
This talk will critically examine these perspectives, discussing whether AI is on the trajectory to achieving human-like cognition or if there are fundamental barriers that limit computational models from replicating consciousness. Ultimately, we will explore whether the rapid scaling of AI is bringing us closer to AGI or revealing the limits of algorithmic intelligence. Is the human brain super-Turing powerful?
This is the fourth talk in the 2025 Data Science Platform Seminar Series.
Speaker’s Bio
John Abela is an Associate Professor in the Department of Computer 福利在线免费 Systems at the
Faculty of ICT. He holds a BSc in Mathematics and Computing (Malta), an MSc in Computer Science (UNB) and a PhD in Theoretical Machine Learning (UNB). His main areas of specialization are AI, machine learning, deep learning, machine vision, optimization, computational complexity, and Large Language Models (LLMs).
Title: From Tradition to Intelligence: RoadEye and the Future of Urban Traffic Management
Speaker: Dr Kenneth Scerri
Date & Time: Wednesday, 19 November 2025, 12:00 (noon)
Location: Faculty of ICT, ICT Communications Lab (Level 0, Block B, Room 1)
Abstract
Urban traffic management has traditionally relied on static models, manual data collection, and rule-based control systems to guide development and operational decisions. However, as cities face increasing congestion and dynamic transport demands, there is a growing shift toward intelligent, data-driven solutions. This talk will review conventional approaches to urban traffic management, outline the state-of-the-art in real-time monitoring and predictive modelling, and discuss future directions in smart transport infrastructure. We will also present RoadEye, a system developed to enhance the Greenroads Ltd. front end solution with real-time data fusion and predictive capabilities. Developed in collaboration with Infrastructure Malta (IM), the project aims to prototype and deploy this solution at future IM developments, supporting more adaptive, efficient, and sustainable traffic management.
This is the fifth talk in the 2025 Data Science Platform Seminar Series.
Speaker’s Bio
Dr Kenneth Scerri is a Senior Lecturer in the Department of Systems and Control Engineering at the University of Malta. His research focuses on data-driven systems engineering, integrating computational modelling, control theory, and artificial intelligence to support decision-making in complex engineering systems. He leads the Intelligent Transport Laboratory, which develops analytical and predictive tools for urban mobility, sustainability, and infrastructure planning. Dr Scerri is the Principal Investigator of RoadEye, an AI-driven platform developed with Greenroads Ltd and supported by Infrastructure Malta to enable real-time data fusion and predictive traffic management. Beyond transport, his work spans biomedical signal analysis and energy system optimisation, supported by multiple national and international grants.
Title: Retrieval-Augmented Generation (RAG): Advances, Challenges and Business Insights
Speaker: Karl Cini
Date & Time: Wednesday, 11 February 2026, 12:00 (noon)
Location: Faculty of ICT, ICT Communications Lab (Level 0, Block B, Room 1)
Abstract
Retrieval-Augmented Generation (RAG) combines large language models with external retrieval mechanisms to ground generated outputs in relevant, up-to-date sources, improving factual accuracy, currency, and explainability. This talk traces the technical origins of RAG and its core architectural building blocks, and surveys current production-grade implementations, including dense, sparse, and hybrid retrieval approaches, vector databases, and reranking strategies. It critically examines the practical limitations and risks of RAG systems, such as retrieval errors, latency, cost, and governance concerns, and explores how the emergence of very large LLM context windows is reshaping the design tradeoffs between retrieval-based and context-heavy approaches. The session concludes with a pragmatic business decision framework, addressing cost, compliance, and performance metrics, to help organisations determine when and how RAG should be adopted within enterprise AI systems. Concrete evaluation techniques and deployment considerations are discussed throughout.
This is the sixth talk in the 2025/6 Data Science Platform Seminar Series.
Speaker’s Bio
Karl Cini is a Research Support Officer within the Department of Artificial Intelligence at the Faculty of ICT, University of Malta, and a Fellow of the Association of Chartered Certified Accountants (FCCA) with over 25 years of experience at the intersection of data, finance, and business strategy. He holds a Master of Science in Data Science from the University of Malta, where he developed advanced expertise in AI integration, data analytics, and data modeling to support data-driven decision-making. Prior to specialising in data science, he built a strong commercial and advisory career as an accountant and business consultant, working with both local and international clients. Alongside his academic role, Karl advises organisations on the practical and responsible adoption of AI and data-driven solutions, supporting strategy definition, analytics-led decision making, and organisational readiness for digital transformation.
Title: Agentomics: Autonomous Machine Learning Experimentation with LLM Agents
Speaker: Andrea Gariboldi
Date & Time: Wednesday, 11 March 2026, 12:00 (noon)
Location: Faculty of ICT, ICT Communications Lab (Level 0, Block B, Room 1)
Abstract
Advances in biomedical research increasingly rely on Machine Learning (ML) to extract insights from growing biological datasets. However, developing and evaluating effective ML models requires significant engineering expertise, creating a gap between modern ML methods and their practical use by experimental biologists. Recent advances in Large Language Models (LLMs) have introduced the possibility of autonomous AI agents that can assist in scientific workflows, including the design and implementation of ML pipelines.
In this talk, we explore the emerging paradigm of agentic systems for scientific ML. We present Agentomics, an autonomous LLM-powered system designed to perform end-to-end ML experimentation for biomedical datasets. Given a dataset, Agentomics iteratively explores modelling strategies, generates training pipelines, evaluates models, and produces reusable artifacts for inference and retraining. The system has been evaluated across multiple biomedical domains, including Protein Engineering, Drug Discovery, and Regulatory Genomics. In several benchmark datasets, Agentomics discovers models that surpass state-of-the-art human-engineered solutions, demonstrating the potential of agentic systems to automate complex ML experimentation.
This is the seventh talk in the 2025/6 Data Science Platform Seminar Series.
Speaker’s Bio
Andrea Gariboldi is a Research Support Officer at the University of Malta and a member of the Bioinformatics for Genomics in Malta (BioGeMT) research team. His work focuses on the development of Large Language Model agents for automating supervised machine learning pipelines. His research also explores scalable database architectures for the efficient storage, management, and retrieval of genomic variant data, supporting the development of robust computational infrastructure for genomics research.
Title: LLMs That Lie Less: Engineering Responsible Language Models for Production
Speaker: Natalia Mallia
Date & Time: Wednesday, 29 April 2026, 12:00 (noon)
Location: Faculty of ICT, ICT Communications Lab (Level 0, Block B, Room 1)
Abstract
Large Language Models are increasingly being deployed in high-stakes applications, yet most discourse around their risks remains at the policy and philosophy level. This talk takes a different approach: responsible AI as an engineering discipline.
Drawing on hands-on experience deploying LLMs in a production EdTech environment, this talk walks through the concrete techniques that separate a convincing demo from a system you'd actually trust with real users. We cover retrieval-augmented generation and the subtler problem of source adherence by getting models not just to access correct information, but to actually use it. We examine guardrail architectures, their latency and accuracy tradeoffs, and the surprisingly non-trivial question of who decides what a "safe" output looks like in context. We look at evaluation strategies that go beyond standard benchmarks and actually measure what matters: factual consistency, grounding, and failuremodes that only surface in the wild. Finally, we discuss cost-reliability tradeoffs: How mixture-of-experts architectures and knowledge distillation let you build systems that are both responsible and economically viable.
This is the eighth talk in the 2025/6 Data Science Platform Seminar Series.
Speaker’s Bio
Natalia Mallia is a Malta-based AI engineer and thought leader specialising in ethical, production grade artificial intelligence. Her work spans the design and deployment of LLM systems across EdTech, manufacturing, and creative industries, with a focus on responsible AI, production reliability, and cost-efficient architectures. She has spoken at NVIDIA GTC, SiGMA, and the World AI Cannes Festival, and was recognised at Malta's Best Businesswoman Awards in 2025.
Title: Transformers 101: How large language models work under the hood
Speaker: Marc Tanti
Date & Time: Wednesday, 13 May 2026, 12:00 (noon)
Location: Faculty of ICT, ICT Communications Lab (Level 0, Block B, Room 1)
Abstract
The last DSP talk of the scholastic year will be an introductory lecture about neural network transformers: how they work, how they are trained, and some different ways they can be used. The lecture will be aimed at students but anyone who is interested is welcome to join.
This is the ninth and final talk in the 2025/6 Data Science Platform Seminar Series.
Speaker’s Bio
Dr Marc Tanti is a computational linguistics lecturer with the Institute of Linguistics and Language
Technology and an associate of the Department of AI. His research interests are in neural networks and natural language processing, especially for low resource languages and the Maltese language.
Title: A simple ZKP scheme of remote human presence based on biometric cryptosystem
Speaker: Norman Poh
Date & Time: Tuesday, 25 February 2025, 12:00 (noon)
Location: Faculty of ICT, ICT Communications Lab (Level 0, Block B, Room 1)
Abstract
A biometric cryptosystem is a family of privacy-preserving biometric algorithms that is designed to work with cryptographic solutions that forms the basis of security in the Internet. Trust Stamp has recently extended this solution, known as stable IT2, which is capable of extracting 256 bits of entropy from an individual's live face capture session. The extracted "stable key" can be bound to any token or secret, which can be used to represent their online presence, their digital identity, their digital travel credentials, single-sign-on sessions, and more.
This biometric binding means that cryptographic secrets no longer need to be strictly stored on a hardware secure module (which is fundamental to the FIDO2 protocols). Since devices can be easily lost or stolen, attackers often exploit this loophole by impersonating the target victim, which often relies on less reliable authentication mechanisms such as recovery emails and passwords. Effectively, stable IT2 replaces this hardware requirement, making it possible to perform account recovery using biometrics, the very same technology that is used for device authentication. As a result, a much higher authentication assurance, at the levels of NIST AAL2 and AAL3, can still be achieved. Therefore, our solution can complement existing FIDO2 protocols and is particularly useful for account recovery and other situations where authentication can only take place without any device. Since the 256-bit secret can only be generated by a live, authenticated presence of the account holder, the solution can ensure a “trusted presence ®”, meaning that a relying party can be sure that only the right person with the right credentials, using the right device, is present right now as they process their transactions.
Speaker’s Bio
Norman Poh is a technical leader at Truststamp, specializing in privacy-preserving biometrics and identity management to address KYC/AML issues. With a diverse background in data science, he has worked in financial forecasting for the oil and gas industry and in healthcare, focusing on disease progression modeling. He is currently an Affiliate Associate Professor at the University of Malta and has previously held various academic positions at the University of Surrey, where he led a Medical Research Council-funded project on chronic kidney disease. Poh has over ten patents and more than 100 peer-reviewed publications, earning multiple awards, including Researcher of the Year at the University of Surrey in 2011. He obtained his Ph.D. in information fusion from EPFL in 2006. Additionally, he has served as an adviser and reviewer for several academic and professional organizations in the fields of biometrics and security. Connect with him at LinkedIn:
Title: Version Control on the Middle-Earth cluster (and elsewhere)
Speaker: Johann A. Briffa
Date & Time: Wednesday, 9 April 2025, 12:00 (noon)
Location: Faculty of ICT, ICT Communications Lab (Level 0, Block B, Room 1)
Abstract
In this third seminar related to the use of the Middle-Earth cluster, we focus on the use of version control to collaborate and to keep track of changes in code and documents. The main focus is on the use of git, through our local GitLab instance, but the same principles apply with other services (such as Github) and other version control systems. We start with an introduction to version control, go through the process of setting up an account and a project repository, and work through the typical day to day interaction with the version control system. We conclude by looking at how to deal with some less frequent situations.
Speaker’s Bio
Johann A. Briffa is a Professor in the Department of Communications and Computer Engineering at the University of Malta. He obtained his PhD in Systems Engineering from Oakland University, Rochester MI in December 2003. His research interests include information theory and image processing, most recently applied to quantum key distribution and remote sensing respectively.
Title: LLM-powered Agents
Speaker: Vlastimil Martinek
Date & Time: Wednesday, 14 May 2025, 12:00 (noon)
Location: Faculty of ICT, ICT Communications Lab (Level 0, Block B, Room 1)
Abstract
Large language models provide us with a very general reasoning tool that can accept instructions and data in natural language format. In recent years, LLMs crossed a threshold where they can be used for task automation. In this talk, we will cover the basics of how LLM Agents work, their advantages and disadvantages, and how we can use them to automate tasks previously unautomatable by traditional programming.
Speaker’s Bio
Vlastimil is a computer scientist working in the field of computational biology. His research focuses on the RNA biology of aging and usage of machine learning in extracting knowledge from large unstructured biological datasets.
Title: An Introduction to QUANGO: QUANtum and 5G cOmmunication
Speakers:
Dr Sergio Aguilar - R&D Engineer, Sateliot
Prof. Francesco Vedovato - Professor, University of Padova
Prof. André Xuereb - Professor, University of Malta
Date: Wednesday, 8 November 2023
Abstract:
Communicating sensitive information across long distances is an essential part of today’s world. However, the current telecommunication infrastructure is open to compromise by quantum computers when these become large enough to be cryptographically relevant. While quantum mechanics poses the challenge, it may also offer the solution. Quantum Key Distribution (QKD) is a protocol based on the laws of quantum mechanics that guarantees information theoretical security in the sharing of cryptographic keys to be used for secure communication.
The implementation of this protocol is at the centre of project Quango’s proposal. Quango, which stands for cubesat for QUANtum and 5G cOmmunication, is an EU Horizon 2020-funded project. Its main objectives include the design of a network of 12U-CubeSat low-earth-orbit satellites offering combined capabilities for communication secured by Quantum Key Distribution (QKD) and for 5G connection for Internet of Things (IOT), the development of payloads, sub-systems, and ground stations for such a network and its implementation feasibility.
During this Seminar, we will be offering an introduction to project QUANGO together with a discussion on various aspects of the project such as the underlying concepts behind QKD and its implementation.
Title: Deep learning insights into messenger RNA decay mechanisms
Speaker:
David 膶echák is a PhD candidate at the Central European Institute of Technology, where he is developing deep learning methods to understand biological data, focusing on messenger RNA decay and microRNAs.
Date: Wednesday, 6 December 2023
Abstract:
RNA decay is a biological process in which RNA molecules are selectively degraded. This degradation is a crucial mechanism for regulating gene expression, maintaining cellular function, and controlling the quality of RNA in the cell. RNA decay is involved in numerous cellular processes and disruptions in these processes can lead to various diseases.
Prediction of RNA decay sites and possibly even decay rates can provide insights into the lifecycle and regulation of RNA, as well as help in understanding diseases that are associated with RNA malfunction. It can also be helpful in designing therapeutic strategies for these diseases.
Title: MARVEL: Audio-Visual Machine Learning Models and Applications in Transport
Speaker:
Adrian Muscat, a Professor at the University of Malta, holds a Ph.D. in Electronics Engineering. His diverse expertise includes computational electromagnetics, optimization, and simulation methods. Engaged also in Research and Development at Greenroads Ltd, he focuses on machine learning, computer vision, and automation for applications in robotics, assisted living, and video analytics.
Date: Wednesday, 20 March 2024
Abstract:
Smart city environments generate large amounts of data from multimodal sources such as video cameras and microphones installed across the cities. Most of this data is largely underutilised and eventually deleted, mainly because of engineering and technology limitations. In an attempt to narrow the gap, MARVEL, a EU Horizon 2020 RIA project, developed an experimental framework to manage the flow and processing of multimodal data over an Edge-to-Fog-to-Cloud (E2F2C) infrastructure, which would allow the end-user (e.g., researchers, engineers, managers or policy-makers) to extract useful information from the raw data via a graphical user interface (GUI). The platform has been demonstrated across three experimental pilots carried out in Trento (Italy), Malta and the University of Novi Sad (Serbia). The use cases are designed from a user-centric perspective and address a number of societal challenges in urban mobility and personal security. Following an overview of the MARVEL framework, we will discuss some of the AI models implemented and their evaluation in the wild. We will conclude the talk with a discussion on the framework’s perceived impact on society.
Title: Sparsity, High-Dimensional Data, Dimension Reduction and the Time Series Setting
Speaker:
Prof. David Suda, University of Malta
Date: Wednesday, 17 April 2024
Abstract:
For the benefit of audiences who may be unfamiliar with the concepts, terms such as sparsity, high dimensional data and dimension reduction will be introduced. Sparsity is the statistical practice of reducing the number of non-zero parameters in a model, a variable selection approach that can commonly be achieved by adding a penalty to the objective function of an estimation problem. High dimensionality, on the other hand, refers to datasets where the number of variables is close to or larger than the sample size. Examples of situations where this can occur are econometrics, imaging (e.g. fMRI data) and genomics. Dimension reduction, on the other hand, is a more commonly known concept – it is the practice of explaining a multivariate setting in lower dimensions, and in classical statistics is most commonly achieved via principal components analysis (PCA). We start by providing a brief overview of the sparsity treatment on classic statistical models – particularly with the intent of handling variable selection in the high dimensional context.
The main focus of this talk will be dynamic principal components analysis (DPCA) which refers to an extension of PCA in a time series setting. The popular dimension reduction technique of PCA needs no introduction with many practitioners, but its lesser-known relative, DPCA, addresses the handling of time dependence and/or short-term correlation not catered for by PCA. Brillinger's frequency domain approach is the earliest of such approaches and is aimed at a single realisation setting. In the last decade, time-domain approaches have also evolved, addressing both the single realisation and multiple realisation settings. Some sparsity extensions for the high-dimensional data setting have also been introduced. Peer-reviewed literature addressing high-dimensionality in the frequency domain setting remains missing.
The frequency domain approach to principal components essentially replicates the classical approach but on cross-spectra instead of the covariance matrix. Alternatively, we can also consider spectral decompositions of the data matrix/Fourier transforms instead of the data matrix itself. From the frequency domain setting, the loadings in the time series domain can then be recuperated through the Fourier inverse, and the principal components through the dynamic Karhunen-Loeve expansion. Our current research aims to address the void concerning high-dimensionality in academic literature when it comes to frequency-domain principal components. Taking a cue from a plethora of literature in the context of sparse PCA, we find that many techniques for addressing sparsity in static sparse PCA setting can be extrapolated to the frequency-domain approach. A way forward is devised regarding its practical implementation, with the long-term aim of also implementing these techniques on real applications.
Title: ADACE3 - Leveraging LLMs for receipt analysis
Date: Wednesday, 15 May 2024
Speaker:
Xandru Mifsud possesses a B.Sc. in Mathematics and Computer Science, coupled with experience in the logistics industry. He is currently an RSO I and Master's student at the University of Malta, forming one of many team members on the ADACE3 project.
Abstract:
Even in today's digital landscape, manual paperwork in accounting tasks still remains prevalent. This presents a significant hurdle to automation, impacting process efficiency and costs. Commercial solutions often offer limited functionality, requiring frequent human intervention or highly skilled personnel. This means that many small businesses cannot afford such solutions, and therefore rely on manual workflows.
The Automated Document Analysis and Classification for Enhanced Enterprise Efficiency (ADACE3) project, funded by the Malta Council for Science and Technology (MCST), aims to address these challenges. By harnessing cutting-edge artificial intelligence algorithms such as self-supervised learning and joint vision-language models, specifically LayoutLM, the project seeks to develop a system that significantly reduces human intervention and is more accessible to small businesses in terms of cost and usability.
Title: ChatGPT for the Language of Life
Speaker: Dr Petr Simecek
Date and time: Wednesday 19 April 2023, 12:00 noon
Venue:
Abstract: The realm of natural language processing (NLP) has witnessed a revolution with the advent of massive language models, such as GPT3.5, OPT, and BLOOM. Recently, similar neural network architectures have been adapted to genomics and proteomics, paving the way for advancements in these domains. In this presentation, we will discuss existing DNA and protein language models, namely DNABert, ProtBertBFD, and ESM2, and illustrate how they can be tuned to specific objectives. Furthermore, we will elucidate how the model's embeddings encapsulate both evolutionary and functional information, highlighting their significance. To conclude, we will demonstrate this methodology by addressing the problem of detecting a topological knot on the protein backbone. Precisely, we will classify proteins to be knotted or not based solely on their sequence.
Title: Quantum communications projects at UM
Speaker: Prof. Andre Xuereb
Date and time: Wednesday 15 March 2023, 12:00 noon
Venue:
Abstract: The security of our digital systems is founded on assumptions that are now known to be flimsy. We will soon have to overhaul all our cryptographic methods and even introduce entirely new technologies, based on quantum mechanics, to solve this problem. The University of Malta is engaged in a number of EU-funded projects that aim to develop the basis of future quantum-secured communication networks, deploy a network in Malta, and even prepare for quantum satellites. In this talk I will explain the foundational principles behind the so-called quantum threat, how quantum mechanics (and some fancy mathematics) may have the solution, and briefly touch upon four projects that we are involved in. Most of all, I want to discuss fully funded opportunities we have for postgraduate positions and studies. We want you to work with us!
Title: Vogt, Bailey and BOB: an exploration of local connectivity in the brain
Speaker: Dr Claude Bajada
Date and time: Wednesday 22 February 2023, 12:00 noon
Venue:
Abstract: My talk will discuss the work conducted at the University of Malta's “Boundaries of the Brain Lab” lab that aims to understand local connectivity in the brain using Magnetic Resonance Imaging (MRI) techniques.
Our group has developed software, the Vogt-Bailey toolbox, which utilises spectral graph theory to objectively measure the degree of homogeneity in cortical neighbourhoods and address the criticisms by Percival Bailey and Gerhardt von Bonin about the limitations of traditional cortical parcellation championed by Oskar and Cecile Vogt (ie. splitting the brain into distinct regions).
I will discuss the inspiration behind the project, discuss the advancements in the field on brain connectivity and the impact of this work on our understanding of the brain.
Title: Data Storage on the Middle-Earth cluster
Speaker: Prof. Johann Briffa
Date and time: Wednesday 7 December 2022, 12:00 noon
Venue:
Abstract: In this second seminar in the series on submitting jobs on the Middle Earth cluster, we consider the issue of data storage. We start with a few details on the different levels of storage available on the cluster, how this impacts our processing pipeline, and the effect of storage choice on speed of execution as well as storage efficiency and scalability. We then look at methodologies to determine what should be stored where, how to express this in a job submission script. A number of common use cases are considered, and the best practice approach detailed for each case.
Title: Pushing the limits of RL in linear environments
Speaker: Dr Leander Grech
Date and time: Wednesday 12 October 2022, 12:00 noon
Venue:
Abstract: The LHC at CERN requires beam-based feedback systems to ensure correct operation. Traditionally, a Proportional-Integral (PI) controller together with a linear model of the beam-based system is used to apply corrections to the superconducting magnets in order to control specific beam and machine parameters, e.g. tune and orbit. Previous work developed a simulation environment for the Tune Feedback (QFB) system and it was shown that an RL agent can outperform a PI controller in cases where classical control algorithms generally fail. In this work, the ideas from this simulation were extended to create an environment called random environment (RE), which represents feedback systems of the same type with the additional functionality that the input and output dimensions can be set arbitrarily and the actions can be either continuous or discrete. The limitations of state-of-the-art RL algorithms are assessed with different configurations of RE, showing for example that popular deep reinforcement learning algorithms such as proximal policy optimization (PPO) perform poorly and unreliably on higher dimensional tasks. Non-parametric methods, which offer some theoretical guarantees, were used to assess the interplay between exploitation and exploration, and also to shed some light on the best practices to follow when training beam-based controller systems.
Title: Efficient use of the Middle-Earth cluster
Speaker: Prof. Johann A. Briffa
Date and time: Wednesday 18 May 2022, 12:00 noon
Venue:
Abstract:
In this seminar we go beyond the basics of submitting a job with the Slurm scheduler, considering various aspects to ensure our jobs run to completion as quickly as possible. We start with a few details on how the scheduler works, how this impacts our job specifications, and how our jobs affect others. We look at methodologies to determine what resources are jobs really need, so that we request the minimum necessary resources. We also consider what other jobs are currently running or in queue, and how to find out what resources are immediately available. Finally, we also look at ways to debug problems with our job submissions.
Title: Disaggregation and Placement of In-Network Programs
Speaker: Dr Nik Sultana
Date and time: Wednesday 23 February 2022, 12:00 noon
Venue:
Abstract: Programmable network switches and NICs are enabling the execution of increasingly rich computations inside the network using languages like P4. Today's in-network programming approach maps a whole P4 program to a single target, limiting a P4 program's performance and functionality to what a single target device can offer. Disaggregating a single P4 program into subprograms that execute across different targets can improve performance, utilization, and cost. But doing this manually is tedious, error-prone and must be repeated as topologies or hardware resources change.
This talk describes Flightplan: a target-agnostic, programming toolchain that helps with
splitting a P4 program into a set of cooperating P4 programs and maps them to run as a
distributed system formed of several, possibly heterogeneous targets.
The talk will cover both systems' and programming language aspects of this research. We will look at evaluation results from testbed experiments and simulation. During the talk I will also describe how Flightplan's design addresses practical concerns, including the provision of a distributed diagnostics interface and the mitigation of partial failures.
Code, documentation, tests, a demo, and videos can be obtained from
Title: Docker: How to easily run an entire software stack locally and ease software
distribution
Speaker: Dr Noel Farrugia
Date and time: Wednesday 19 December 2021, 13:00
Venue:
Abstract:
Most of us researchers have no doubt dealt with either libraries or software that requires a number of steps and dependencies to be installed before they can be used. This process is rarely as easy as the user manuals make it to be. Docker can be the solution to this both as a consumer and producer of software. Docker gives you the ability to ensure that users of your software are running an identical setup to that specified by you in the docker file easing the barrier of entry to the use of such software.
In this seminar we will discuss what are docker containers, what is the difference between containers and virtual machines, how to create your own docker container and more.
Title: The VBIndex Toolbox: studying correlations in human brain function using fMRI data
Speaker: Dr Christine Farrugia
Date and time: Wednesday 15 December 2021, 12:00 noon
Venue:
Abstract: The development of Magnetic Resonance Imaging (MRI) has been instrumental to our understanding of the function and structure of the human brain. In this talk, we start by taking a look at the main features of this imaging technique and the associated pre- and post-processing pipelines. We then move on to the VBToolbox, a functional MRI (fMRI) analysis software package developed by the Boundaries of the Brain (BOB) group at the University of Malta. This package makes use of principles from spectral graph theory to detect correlations in the function of the human brain, both locally (at each voxel) and also across the whole brain. The data series collected during an MRI scan session are each associated with a volume element (voxel) within the brain, and the main idea is to consider each voxel as the node of a graph whose edges are weighted by the degree of similarity between the different series. We shall review some of the results obtained with the VBToolbox and discuss their implications, and end the talk by looking at current developments and future work.
Title: The Deep-FIR Project: Super-Resolution in the Wild
Speaker: Mr Matthew Aquilina
Date and time: Wednesday 10 November 2021, 12:00 noon
Venue:
Abstract: Super-resolution (SR) involves enlarging and enhancing low-resolution images using computational techniques to accurately predict the unseen details of an image (‘Zoom and Enhance’). The influx of deep learning has pushed SR performance well beyond what was previously thought possible with mathematical/statistical techniques alone. With today’s convolutional neural networks (CNNs), we can super-resolve artificially degraded (blurred, downsized, etc.) images into incredibly detailed and realistic results, or near-exact replicas of their original high-resolution source. However, SR is far from being a solved problem, as applying those same networks on real in-the-wild images (taken by a CCTV camera, smartphone, etc.), results in significantly worse quality images, sometimes no better than the standard digital zoom available on our smartphones. In-the-wild images pass through a large (and typically unknown) number of degrading operations (lens blurring, image compression, noise influx, etc.) before they are eventually saved onto our devices. Even the largest neural networks are unable to decipher and reverse such a complex web of degradations without guidance. In the Deep-FIR project, our aim is to introduce new methods for equipping neural networks with the tools to identify and combat various degradations from any input image or video.
Our first step in this direction has been to introduce meta-attention, a lightweight mechanism which allows users to exploit image metadata (attributes) to configure any SR network to identify and counteract various degradations.
This seminar will delve into the details of meta-attention, as well as highlight and discuss the various research fronts the Deep-FIR project is investigating across both image and video SR.
Title: LigityScore: Convolutional Neural Network for Binding-affinity Predictions
Speaker: Mr Joseph Azzopardi
Date and time: Wednesday 23 June, 12:00 noon
Venue:
Abstract: Scoring functions are at the heart of structure-based drug design and are used to estimate the binding of ligands to a target. Seeking a scoring function that can accurately predict the binding affinity is key for successful virtual screening methods. Deep learning approaches have recently seen a rise in popularity as a means to improve the scoring function having as a key advantage the automatic extraction of features and the creation of a complex representation without feature engineering and expert knowledge. In this seminar we will present LigityScore1D and LigityScore3D, which are rotationally invariant scoring functions based on convolutional neural networks. LigityScore descriptors are extracted directly from the structural and interacting properties of the protein-ligand complex which are input to a CNN for automatic feature extraction and binding affinity prediction. This representation uses the spatial distribution of Pharmacophoric Interaction Points, derived from interaction features from the protein-ligand complex based on pharmacophoric features conformant to specific family types and distance thresholds. The data representation component and the CNN architecture, together, constitute the LigityScore scoring function. The main contribution of this study is to present a novel protein-ligand representation for use as a CNN-based SF for binding affinity prediction. LigityScore models are evaluated for scoring power on the latest two CASF benchmarks. The Pearson Correlation Coefficient and the standard deviation in linear regression were used to compare and rank LigityScore with the benchmark model, and to other models recently published in literature. LigityScore3D has achieved better overall results and showed similar performance in both CASF benchmarks. LigityScore3D ranked 5th place for the CASF-2013 benchmark, and 8th for CASF-2016, with an average R-score performance of 0.713 and 0.725 respectively. LigityScore1D ranked 8th place for the CASF-2013 and 7th place for CASF-2016 with an R-score performance of 0.635 and 0.741 respectively. Our methods show relatively good performance when compared to the Pafnucy model (one of the best performing CNN-based scoring functions), on the CASF-2013 benchmark using a less computationally complex model that can be trained 16 times faster.
Title: Machine Learning for Particle Accelerators
Speaker: Dr Ing Gianluca Valentino
Date and time: Wednesday 3 March, 12:00 noon
Venue:
Abstract: Although machine learning techniques have been applied to particle accelerators since the late 1980s, a renaissance has only been seen in recent years. This is due, in part, to the success of modern developments such as deep learning and, in part, is a result of the
sophistication and data-intensiveness of current machines. The system dynamics of particle
accelerators tend to involve large parameter spaces which evolve over multiple time scales,
and interrelations between accelerator subsystems may be complex and nonlinear.
As a result, there is growing interest from the particle accelerator community to use machine learning techniques to analyze large quantities of archived data to accurately model accelerator systems, detect anomalous machine behavior, and perform active tuning and control. It is expected that machine learning will become an increasingly valuable tool to meet new demands for beam energy, brightness, reliability, and stability. This seminar will review the ongoing research activities in this area, as well as the contribution of the University of Malta in collaboration with the particle accelerator community.
Title: Stereo Vision from Earth Observation
Speaker: Dr Mang Chen
Date and time: Wednesday 17 February, 12:00 noon
Venue:
Abstract: The number of Earth observation satellites has increased drastically over the past decade, where some of these satellites enable the capture of two (or more) images of the same region at quasi real-time. Stereo vision techniques can then be used to automatically
compute Digital Elevation Models (DEMs) that are important for a number of domains including hydrology, urban planning and natural hazard detection. These stereoscopically derived DEMs provide an efficient and low-cost means for remote mapping of surface topography over large areas and at multiple times for change detection.
The Centre National d’Etudes Spatiales (CNES) have developed the Stereo Pipeline for
Pushbroom Images (S2P) framework that combines the information obtained from the
Satellite together with a stereo matching process to estimate the DEM. However, the stereo matching process adopted by this framework is based on classical techniques. The aim of the SAtellite TraIning and NETworking (SATINET) project is to adopt deep-learning based techniques to improve the stereo-matching process. WorldView-3 satellite images at a resolution 30cm and airborne Lidar data covering the area of San Fernando in Argentina was adopted in our evaluation. Compared with the inherent shortages of classical techniques in feature extraction for textureless, repeated pattern and occlusion, deep learning methods automatically learn and calculate feature parameters through training. Compared with the 66.85% completeness of the classical techniques (SGBM), our method reaches 74.05%, which is a 7% gain. The results below further show that our approach is more robust when compared to the state-of-the-art method.
Title: Can we connect Vision and Language using Graphs?
Speaker: Mr Brandon Birmingham
Date and time: Wednesday 16 December, 12:00 noon
Venue:
Abstract: A long-standing goal of Artificial Intelligence is to have agents capable of understanding and interpreting the visual world using natural language. The advancements in computing power and the sheer amount of visual and linguistic data available today helps in getting closer to this quest. Research at the intersection of Computer Vision and Natural Language Processing is currently booming and the automatic generation of image captions has recently gained a lot of popularity. Several ideas and architectures have been proposed to machine-generate human-like sentences that describe images, but all are short of reaching human-level quality. The focus of this talk is to specifically explore how the graph data structure can be used to connect the vision and language modalities in the context of image caption generation and how such graph-based models compare with the current state-of-the-art deep learning-based models.
Title: Enhancing machine learning with synthetic biometrics and documents
Speaker: Dr Norman Poh
Date and time: Wednesday 11 November, 12:00 noon
Venue:
Abstract: Developing machine learning models requires a lot of data. The more data there is, the lower the risk of overfitting. Traditionally, we have to collect real data. While this is laborious, annotating data is even more so. We resolved this by creating tools to annotate data automatically, leaving only refined annotation to be done by humans (using CVAT). Next, we explored options to generate synthetic data. By mixing real and synthetic data, we have solved a number of computer vision problems effectively, from object detection to classification and semantic segmentation. In this talk, I will share with you our journey in biometrics and document verification.
Title: A gentle introduction to quantum computing
Speaker: Dr John Abela
Date and time: Wednesday 28 October 2020, 12:00 noon
Venue:
Abstract: Quantum computing is an area of study that focuses on the development of computing devices that are based on the principles of quantum mechanics. Quantum mechanics is a theory that attempts to explain the nature and behavior of energy and matter on the microscopic (atomic and subatomic) level. Quantum computers use a combination of Qubits (Quantum bits), and the quantum phenomena superposition and entanglement, to perform specific computational tasks. All this at a much higher efficiency than their classical counterparts. Quantum computers are not super-Turing powerful but they provide an exponential speed-up for certain NP-Hard problems and for specific use cases. Development of quantum computers is progressing at a fast pace with billions of dollars being poured into research and development. Quantum computing has important implications for areas such a cryptography and will completely revolutionize drug discovery and design. In the talk we will introduce the basic ideas of quantum mechanics and then give a brief overview of how quantum computers work.
Title: Data warehousing and analytics
Speaker: Andrew Sammut
Date and time: 19 February, 12:00 noon
Venue: Communications Lab, Faculty of ICT (Room 1, Level 0, Block B)
Abstract: Building a data warehouse and maintaining it is not an easy feat. In order to design a proper data warehouse, one has to understand the data underlying the structure, as well as the question or problem that needs to be answered. When one fails to do so, this can be catastrophic for a business that is heavily dependent on its data. In his presentation Andrew Sammut will be discussing the best practices for designing data warehouses on the cloud, as well as providing solutions for some common challenges. He will demonstrate how DAX measures can be used in PowerBI to analyse and create models using the underlying data. Further to this, he will be demonstrating how AI models can be deployed on the cloud stack to be used for modelling purposes. This includes the comparison of different AI models used for time series forecasting including ARIMA, neural networks and decision trees.
Title: Mummies, tomography, and segmentation: The ASEMI Project
Speaker: Marc Tanti
Date and time: 15 January 2020, 13:00
Venue: Communications Lab, Faculty of ICT (Room 1, Level 0, Block B)
Abstract: Ancient Egypt is known for mummifying pharaohs but did you know that they also
mummified animals? In order to investigate this practice and the reasons behind it,
archeologists at the ESRF use x-ray tomography in order to produce 3D scans of what is
inside these mummies without destroying them. This results in a greyscale volume showing
just the different densities of materials inside but it would be more useful to be able to
recognise and highlight the different objects such as bones, textiles, and biological tissues, a process that takes months of manual work to do. The ASEMI project is a research-based
project to use computer vision techniques to automatically segment these 3D volumes into
different materials which should cut down the time required to analyse these mummies
from a few months to a few days. This talk will go through the basics of animal mummies,
tomography, and segmentation in an accessible way.
Title: A Heuristic Solution for the Selective Dial-a-Ride Problem
Speaker: Mark Cauchi
Date and time: 11 December 2019, 12:00 noon
Venue: Communications Lab, Faculty of ICT (Room 1, Level 0, Block B)
Abstract: Demand responsive transportation can relieve road congestion and pollution, offering reliable transportation at a cheap price. Such a service can be implemented as a mobile application, and requires three stakeholders, namely, the service provider, passengers, and drivers. An algorithm is required to assign passengers to drivers, while delivering some optimal solution based on quality of service and/or profits. This presentation addresses a variant of this problem referred to as Selective Dial-A-Ride Problem (DARP). Solving large instances of this problem may be infeasible in reasonable computational time, and therefore, a metaheuristic solution is often adopted. The Variable Neighbourhood Search (VNS) has emerged as the most prominent modern solver to such problems, with such advantages as control over the global and local searches of the algorithm. In this presentation, a VNS is shown with two algorithmic novelties. Statistical analysis of test results based on a local scenario, yield invaluable information to the service provider which they may share with application users, such as, the expected profitability to the driver and the chances of being served to the passenger. Such information is indispensable for the successful adoption of this application.
Title: Enhancing Satellite Imagery with Oriented Filters and Machine Learning
Speaker: David T. Lloyd
Date and time: 30 October 2019, 12:00 noon
Venue: Communications Lab, Faculty of ICT (Room 1, Level 0, Block B)
Abstract: There are currently over 600 Earth-observing satellites in orbit and that number is set to grow in the years to come. Together these satellites produce many 100’s of TB of data each day, with a significant fraction of this data made available for free to users. Naturally such large datasets are ripe for the application of newly developed tools originating in data science, notably those utilising deep learning. In this talk I will present some preliminary results from the SAT-FIRE project. I will describe a simple, computationally-efficient filter we have developed for improving the quality and fidelity of satellite images acquired with poor signal to noise ratio. Further, I will show that computational super-resolution techniques based upon neural networks are able to enhance the resolution (and information content) of thermal images of the sea around the Maltese islands. Such enhanced images are of use for improving the accuracy of oceanographic models, with applications in climate modeling and directing sea rescue efforts.
Title: Machine Learning in Computer-Aided Drug Discovery (Workshop)
Speaker: Dr Jean-Paul Ebejer
Date and time: 24 May 2019, 12:00 noon
Venue: Networks Lab, Level -1 block B, Room 7, Faculty of ICT
Abstract: Computer-Aided Drug Design (CADD) plays an increasingly critical role in the drug-discovery process. CADD involves the application of computer algorithms to improve pharmaceutical productivity. These include algorithms for the identification of the biological target involved in a disease, toxicity and side-effect prediction, and searching a database for molecules which exhibit a therapeutic effect against a particular protein of interest. The latter is known as Virtual Screening. In this workshop I will give an overview of CADD with particular emphasis on virtual screening. We will develop a machine learning (ML) model to discriminate between actives and decoys against a protein target which plays a critical role in the life-cycle of HIV. This interdisciplinary talk is aimed at an audience with prior Python programming experience and an interest in the application of ML models in life sciences.
