Welcome to the Seminar Series of the School of Computer Science and Information Technology

The minimean complexity is the minimum average time an algorithm can take to solve a given problem. A minimean lower-bound indicates at which point algorithmic redesign no longer can yield a “big-O time” improvement.

For sorting the minimean is the sharp Omega(nlogn) bound: no comparison-based sort executes in less than cnlogn average-case time (for some c > 0) and some sorts achieve this lowest time.

For non-sorting problems Yao’s elegant argument [1] yields a generalised minimean bound. Contrary to the sorting case this bound need not be sharp. It can undershoot the sharp lower bound by an asymptotic order.

A new multi-target bound R is presented for “calibrated refiners”—a class of comparison-based algorithms. The R-bound coincides with Yao’s bound in case of a single target.

The lack of sharpness of Yao’s bound (and its generalisation) is highlighted on examples. This leads to an new interpretation of the bound. A different technique is presented which yields Quicksort-Partitioning’s sharp bound, shedding light on the genius of Tony Hoare’s original design. Finally, ongoing work on deriving sharp bounds is discussed.

* This work was funded in part by a Fulbright scholarship carried out at Stanford University (research host Don Knuth).

[1] On the complexity of partial order productions, Andrew Chi-Chih Yao, SIAM J. COMPUT. Vol. 18, No.. 4, pp. 679 - 689, August 1989. (Lists multi-target bounds as an open problem.)

The verification of software systems often relies on a combination of formal methods, testing and simulation based approaches. Knowing what to specify for verification cannot be achieved without clear, unambiguous descriptions of the system requirements. This makes detailed requirements elicitation and formalisation a crucial step in the process.

In this talk, I will share experiences from case studies where we have used NASA's Formal Requirement Elicitation Tool (FRET) to formalise natural language requirements, enabling the use of mathematically based techniques to guarantee that the system obeys certain properties. These formalised requirements subsequently guide the systems verification using a combination of paradigms, providing for system-level modelling and verification using model checkers and deductive verifiers. I will also discuss ongoing work in the Science Foundation Ireland Frontiers for the Future Programme, working on verification and visualisation of AI based systems (MAIVV, 2021-2025).

Look around, in many technical fields in computing women are in the minority. This wasn’t always the case and we know much of what is wrong but we are not making progress. Why is that and what can we do make a change? What can we leverage to help provide the supports needed to move forward? In this talk Ruth will share insights on the unique challenges faced by women in technology, highlight successful initiatives and programs aimed at promoting female participation, and discuss the role of mentorship and community support in empowering the next generation of women leaders.

We introduce Logical Credal Networks (or LCNs for short) -- an expressive probabilistic logic that generalizes prior formalisms that combine logic and probability. Given imprecise information represented by probability bounds and conditional probability bounds on logic formulas, an LCN specifies a set of probability distributions over all its interpretations. Our approach allows propositional and first-order logic formulas with few restrictions, e.g., without requiring acyclicity. We also define a generalized Markov condition that allows us to identify implicit independence relations between atomic formulas. We propose novel exact and approximate inference algorithms for computing posterior probability bounds on given query formulas as well as for generating most probable (partial) explanations of observed evidence in the network. We evaluate our method on benchmark problems such as random networks, Mastermind games with uncertainty and credit card fraud detection. Our results show that the LCN outperforms existing approaches; its advantage lies in aggregating multiple sources of imprecise information.

Reinforcement Learning (RL) has seen major breakthroughs in the recent years and is extensively investigated in a range of practical applications, including those within city-scale infrastructures. However, existing algorithms still fall short of being suitable for a wider use in such complex environments.

My research focuses on developing techniques that enable the use of RL for optimization in large-scale adaptive systems, for example, communication networks and intelligent mobility. These systems share properties with many other large-scale systems, i.e., are characterized by distributed control, heterogeneity, presence of multiple and often conflicting goals, reliance on diverse sources of information, and the need for continuous adaptation.

In this talk I will discuss a range of techniques we have developed for enabling RL use in such environments, such as multi-agent multi-objective optimization, state space adaptation in non-stationary conditions, and online transfer learning. In particular, I will focus on explainability of RL systems (XRL), as a crucial element required for ensuring trustworthiness of RL-based systems. I will discuss differences required in explaining RL systems compared to other types of XAI, and present our approaches to generating and evaluating counterfactual and semi-factual RL explanations.

In this seminar, I will discuss the importance of capturing the fine motor skills associated with handwriting. Fine motor skills, i.e. the precise movements we make with our fingers and hands, are crucial for manoeuvring a pencil or pen in a writing task.

Unfortunately, there are fewer opportunities to use or practice handwriting, especially as technology permeates our everyday lives. This shift has sparked interest in preserving handwriting and associated movements. Motion capture technology, commonly used in clinical settings, offers a promising avenue for recording these movements. Nonetheless, more research is needed to understand the user’s experience wearing motion capture tools while their writing is being recorded.

Further in this talk, I will discuss the preliminary analysis of a study where we investigated how two motion capture technologies could be employed to capture fine motor skills connected to everyday handwriting. I will further highlight the preliminary implications these findings could have on how motion capture technology could be designed for this purpose.

Recommender Systems (RS) traditionally prioritize accuracy as their primary evaluation metric. However, this focus on accuracy often comes at the expense of diversity and novelty in recommendations, impacting user satisfaction. Balancing these conflicting objectives necessitates a robust framework. Some studies have leveraged Reinforcement Learning (RL) and Multi-Objective Optimization (MOO) to reconcile these trade-offs, resulting in Multi-Objective Recommender Systems (MORS). However, existing MOO-based MORS solutions may yield suboptimal results, while RL-based approaches may introduce bias and high variance.

In this talk, we present a novel approach to address MORS challenges. We reframe MORS as a sequence modeling problem, introducing MODT4R—a method that predicts user actions based on their sequential trajectory and desired multi-objective returns. MODT4R offers a more effective solution to enhance RS performance while ensuring diversity, novelty, and accuracy in recommendations.

Early efforts to accommodate people with special needs typically took the form of Assistive Technology (AT) - applications designed for specific groups of users with special needs. Some AT solutions proved very successful in meeting the needs of the target users, but use of AT often led users into 'digital ghettos', unable to fully collaborate and share files, etc., with users of mainstream technology. More recently, developments in AT have focussed on adaptations which allow those with special needs to use mainstream software - albeit, in many cases, with a significant reduction in usability. The current trend is towards Universal Design, the aim being to design applications that meet the needs of as wide a range of users as possible without the need for specialist add-ons. However, designing effectively for a wide range of users poses significant challenges, and may not produce solutions that are as effective as those tailored to a specific group of users.

In this talk I will look at some of the tools and guidance available to help those aiming to create inclusive digital solutions, and at some of the challenges we face in attempting to translate theoretical principles of accessibility into effective practice.

AI is fast becoming a significant consumer of the world’s computational power, so it is crucial to use that power wisely and efficiently. Our approaches to doing so must span all levels of the research stack: from fundamental theoretical understanding of the loss surfaces and regularization properties of machine learning models, to efficient layout at the transistor level of floating-point multipliers and RAM. I will talk about projects, such as real-time computer vision on the Microsoft HoloLens HPU (about 3.5 GFLOPS ), which required extreme efficiency in both objective and gradient computations, and how this relates to the training of massive AI models on Graphcore’s IPU (about 350 TFLOPS). Key to this work is how we empower programmers to communicate effectively with such hardware, and how we design frameworks and languages to ensure we can put theory into practice.

So this talk contains aspects of: mathematical optimization, automatic differentiation, programming languages, and silicon design. Despite this range of topics, the plan is for it to be accessible and useful to anyone who loves computers.

The European Commission through its recent document ‘Industry 5.0, a transformative vision for Europe’, has highlighted the role that digital technologies will play in enabling more sustainable economic models. A separate EIT report went on to highlight the need to pilot AI in production processes, in areas such as predictive maintenance, particularly for SMEs, which as the report highlighted, are more conservative and risk averse. However, the ability of SMEs to adopt this technology is challenged by knowledge shortages with the need to up-skill considered crucial.

The A3AM Decision Support Platform provides SMEs with a means to pilot the adoption of AI without a need to develop organic artificial intelligence expertise. Advanced forms of analysis are made accessible, while rules can be created to capture in-house manufacturing expertise, allowing it to be reapplied and scaled to larger manufacturing endeavours. The platform successfully facilitates almost real-time feedback on process data to operators, with decision support based on data analysis, as well as the results of previous process runs. This system has been successfully trialled for additive manufacturing processes (laser powder bed fusion), on production scale processes, including those within industry.

In 1955, John McCarthy and colleagues proposed an AI summer research project with the following aim: “An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.” More than six decades later, all of these research topics remain open and actively investigated in the AI community. While AI has made dramatic progress over the last decade in areas such as vision, natural language processing, and robotics, current AI systems still almost entirely lack the ability to form humanlike concepts and abstractions.

In this talk, I will argue that the inability to form conceptual abstractions—and to make abstraction-driven analogies—is a primary source of brittleness in state-of-the-art AI systems, which often struggle in adapting what they have learned to situations outside their training regimes. I will reflect on the role played by analogy-making at all levels of intelligence, and on the prospects for developing AI systems with humanlike abilities for abstraction and analogy.

• How do we partition machine learning algorithms across Edge-Network-Cloud resources (often referred to as the "Cloud-Edge Continuum") based on constraints such as privacy, capacity and resilience?
• Can machine learning algorithms be adapted based on the characteristics of devices on which they are hosted? What does this mean for stability/ convergence vs. performance?

The presentation will cover a range of topics and will be of interest to those interested in data analytics and machine learning, chronobiology, personal data, student wellness, and educational analytics. Our students leave digital footprints all over their social media and their web browsing and their other online and offline (real world) activities. They also leave digital footprints from their online University activities and that offers the opportunity of whether there is anything useful we as a University can do with this. In the past we have used out-of-the box machine learning techniques to use these University-based digital footprints to predict students' module/course outcomes and we've shown for 000's of students in DCU in their first year, how we can help them by alerting them as to their progress and likely outcomes for those high-failure modules. That results in an increased pass rate in DCU and students heading to the US on J1 visas instead of having to repeat modules during their Summertime. If all we are interested in is helping students pass exams that's fine, but can we do more ?

The FLOURISH module at DCU and UCD focus on overall student wellness and wellbeing and is like similar modules taught elsewhere at places like Yale and Berkeley. However I believe that those wellness or science of happiness courses elsewhere miss a trick in that they do not use students' personal data in any way. The syllabus for FLOURISH at DCU covers topics including sleep, nutrition, physical activity, behaviour change, digital footprints, cognitive psychology, healthy choices and more. For most of these there is a personal data aspect where students get to see how their own personal data can be used by them as a force for their own good as they learn about those wellness-related topics. In addition to gathering and using their own personal data for these topics, we also use some of the digital footprints that the University gathers to feed back indicators of their overall periodicity intensity. The overall outcome is that students learn about aspects of their wellness and wellbeing which are not taught elsewhere and which they may mis-learn from unreliable sources such as social media or their peers. Additionally, the personal data aspects of their own digital literacy is enhanced by them seeing their data used for good.