The quality, accessibility, and affordability of healthcare are being threatened by personnel shortages, high administrative burden, and an ever-increasing demand for healthcare services. The recent introduction of large language models (LLMs) has led to enormous excitement within the healthcare community about their application in healthcare. The question arises if LLMs and other forms of NLP can live up to these expectations.

As part of my PhD, I have worked on several NLP projects within the LUMC, with the aim of integrating them in clinical practice. As I’m finalizing my dissertation, I reflect on the recent advancements in relation to the unruliness of healthcare practice, and what we need to go from paper to practice.

There is no recording and no slides available for this seminar.

The ISSC is planning to configure and maintain a safe platform for using Large Language Models at Leiden University, both locally run and through Microsoft Azure. The platform could significantly speed up the adoption of AI-based tools in education, research and operations. This talk will explore how incredibly easy it is to create such a platform, and why it’s also incredibly hard at the same time.

Join us for an insightful presentation by Dr. Niki van Stein on "Explainable AI in Industry," where the burgeoning field of Explainable Artificial Intelligence (XAI) is demystified. This talk explores the pivotal role of XAI in enhancing transparency and trustworthiness in AI applications, particularly in industrial contexts. Dr. Van Stein will delve into the challenges and advancements in XAI, offering a unique perspective on how this technology is reshaping our understanding and utilization of complex AI systems. Don't miss this opportunity to engage with cutting-edge research and discussions on the future of AI in the industry.

Disclaimer

Disclaimer: Note that the abstract was partly generated by generative AI, the generative AI model also provided an explanation on how the abstract was generated.

To create the abstract, I synthesized the key themes and objectives of the presentation from the provided PDF document. The focus was on making it appealing to a scientific audience while summarizing the essence of the talk. I emphasized the importance and relevance of Explainable AI (XAI) in the industry, as highlighted in the presentation, and portrayed Dr. Van Stein as an expert in the field. This approach aimed to generate interest and convey the significance of the topic, illustrating the talk's value in understanding and applying XAI in industrial contexts.

Several scholars (Laquer & Copus 2017; Leibovitch 2017; Chiao 2018) argue that machine-learning algorithms can and ought to be used by judges at the sentencing stage to predict the statistically typical (i.e., modal) sentencing decision taken by other (actual or counterfactual) judges in relevantly similar cases, and adjust their individual sentences to cohere with these latter ones. Unlike currently deployed AI-informed tools, the proposal here is to use algorithms to predict judicial, not offender behavior. Furthermore, unlike existing static actuarial tables or sentencing guidelines and grids, such algorithms proceed dynamically – viz., by updating sentence predictions based on the decisions taken by individual judges. The contention is that these algorithmic tools can secure more consistency among sentencing decisions while preserving judges’ substantive commitment to reasonably defensible penal principles. The argument of this paper is twofold. First, it argues that the decision-making situation that such proposals would instantiate is that descriptively satisfies the conditions of the Condorcet Jury Theorem (CJT) – viz., one where the average competence of decision-makers is better than random, where decision-makers share the same goal, and where their judgments are independent. Because of this, the proposal seems epistemically desirable. Second, I draw on List & Pettit (2004) and Dunn (2018) to further argue that, insofar as they believe that sentencing algorithms create situations that satisfy CJT, judges are justified to adjust their sentencing decisions to statistically typical ones. Insofar as this happens, and because sentencing is a temporally deployed process, judges’ beliefs that CJT is satisfied would also rationally motivate them to free-ride on other judges’ decisions, and thereby eventually prompt a situation that violates the independent judgment condition posited by CJT. Thus, envisaged diachronically, the proposed algorithms are an epistemic liability. Epistemic free-riding, I note, is both different from and rationally more difficult to address than algorithmic complacency (Zerilli 2021) or automation bias (Kazim & Tomlison). Third, I examine three institutional set-ups that could contain the epistemic free-riding, and conclude that none of them is simultaneously feasible and desirable at the level of penal sentencing practice.

References:

1. Boland, P. J. (1989). Majority systems and the Condorcet jury theorem. Journal of the Royal Statistical Society: Series D (The Statistician), 38(3), 181-189
2. Chiao, V. (2018). Predicting proportionality: The case for algorithmic sentencing. Criminal Justice Ethics, 37(3), 238-261
3. Dunn, Jeffrey, 'Epistemic Free Riding', in H. Kristoffer Ahlstrom-Vij, and Jeffrey Dunn (eds), Epistemic Consequentialism, Oxford University Press 
4. Kazim, T., & Tomlinson, J. (2023). Automation Bias and the Principles of Judicial Review. Judicial Review, 28(1), 9-16
5. List, C., Goodin, R.E. (2001). Epistemic Democracy: Generalizing the Condorcet Jury Theorem, Journal of Political Philosophy, 9(3): 277-306
6. List, C., & Pettit, P. (2004). An epistemic free-riding problem?: Christian List and Philip Pettit. In Karl Popper (pp. 138-168). Routledge
7. Zerilli, J. (2021). Algorithmic sentencing: Drawing lessons from human factors research. In J. Ryberg, & J. V. Roberts (Eds.), Sentencing and Artificial Intelligence (pp. 165-183). (Studies in Penal Theory and Philosophy). Oxford University Press

The most prominent debate in ethnographic museums today revolves around questions of ownership and provenance: how and why did objects move and end up in museum collections – and should they remain where they are today? The main issue in this type of research is that it is highly time-consuming as it requires researchers to manually sift through and read archival material in different museums around the world. Using Large Language Models, Knowledge Graphs, and Network Analysis has the potential to make this research much more efficient. These techniques can enable researchers to study these processes of movement and collecting at unprecedentedly large scales. In this short talk, I’ll speak about how I plan to use some of these methodologies to move the scale of ethnographic object provenance research from the local and incidental, to the global and structural. Additionally, I’ll briefly reflect on how we can embed these new methodologies – and the new questions they raise – in established art historical/museological thinking and theory.

Towards an Automatic and Detailed Understanding of Videos

Thanks to revolutions in supervised learning and the emergence of large, labelled video datasets, tremendous progress has been made in video understanding to automatically understand what is happening. However, current algorithms cannot understand more detail such as identifying how actions happen, which is key to achieving the desired outcome. For instance, existing models could recognize someone performing CPR, but fail to identify it needs to be done faster, firmer and further up the body to have a chance of resuscitation. This talk will cover what is currently possible in video understanding as well as several recent works aiming more a more detailed understanding by (1) skill assessment, (2) identifying relevant adverbs and (3) reducing the supervision needed to distinguish subtly different actions.

Humans seem to regard robots as social actors, they show empathy and altruism toward them and are concerned for their wellbeing. But why exactly would this be the case? One hypothesis is confusion: we don’t know exactly what they are, so we treat them as humans as a cognitive heuristic. Another hypothesis is the similarity hypothesis: what looks human is treated as human. In this talk, I will discuss the role of robots’ similarity to humans and its effect on prosocial behaviour, as well as problems surrounding such behaviour in the lab, and it seems that these competing hypotheses do not have a clear winner between them.

There is an increasing center-periphery political divide in many countries, which has become attenuated by burdens of sustainability reform affecting peripheral areas. Frequently, efforts to engage a broad base of citizens relies on dialogue and decision-making through digital platforms that receive public input or broadcast information. Despite the need to establish viable local sustainability policies, limited attention has been paid to the impacts of center-periphery divides on the effectiveness of digital local policy engagements with citizens.

In this project, we raise the following two questions: (1) how do peripheral municipalities organize public engagement for sustainability through digital means, and (2) what impact does the engagement have on local and national sustainability policies? Using a mixed methods approach, we will focus on understanding and explaining digital types of public engagement for sustainability. The project focuses on the Netherlands given the recent Dutch crises including farmers knocking down the town hall doors because of the new Nitrogen rules, and local actions against wind farms.

Whenever you have learned something from data, you have implicitly also compressed the data. In this talk we explore how we can turn this observation around and use compression to learn explainable models from data. To this end we look at how we can model and quantify information in data in an interpretable manner, and how we can use this for data mining and machine learning.

In this SAILS Lunchtime Seminar, Dr Michael Klos discusses the standards set by online platforms in their community guidelines for manipulated (e.g., edited photos and videos) and synthetic (generated) media.

Many of our social insitutions are based on a model of rational agency that goes back to the Enlightenment. The model features a cognitive and moral profile for individuals (a set of rational competencies that everyone should ideally possess), as well as a set of social conditions that must be met for persons to be able to develop and use their rational competencies in an optimal way (e.g., access to information, free exchange of ideas, a private as well as a public sphere, etc.). 

Everyone (at least 100 academic papers) claims to be able to predict court decisions. Many claim accuracies of over 90%. Now GPT-4 passed the bar. The promises (and the alarms) of a robojudge have been around for years, yet nothing seems to actually happen. In this seminar it was discussed why that is. We addressed how these systems are built, who they are for, who is developing them, and at what stage of the process things seem to go wrong. We then discussed what some of the potential solutions to the crisis within the field of legal judgement prediction are. The talk is aimed at anyone interested in legal tech regardless of their background.

Jelena Prokic, Assistant Professor at the faculty of humanities

Ever since the emergence of Artificial Intelligence (AI), the vast majority of research at the intersection of the Humanities and AI was dedicated to the ethical side of AI and how AI impacts our cultures and societies. While philosophy and critical thinking of AI remain indispensable research areas in the Humanities, lately there has been a surge of research that uses AI to investigate diverse questions in the Humanities. In this talk, Jelena Prokic addressed the recently established research group Leiden Humanities AI and NLP group, Leiden HumAN, which works at the intersection of AI and the Humanities. 

Tycho de Graaf, Professor Civil Law

EU Liability for AI

If defective AI systems are sold or put on the market, the seller and manufacturer may be liable for damages caused by such AI system. In this SAILS Lunch, Tycho de Graaf will provide a high-level overview of a number of legislative instruments the EU has enacted/proposed and which are relevant to such liability, more in particular:

1. Directive (EU) 2019/771 of the European Parliament and of the Council of 20 May 2019 on certain aspects concerning contracts for the sale of goods
2. Directive (EU) 2019/770 of the European Parliament and of the Council of 20 May 2019 on certain aspects concerning contracts for the supply of digital content and digital services
3. Proposal for a Revision of the Product Liability Directive
4. Proposal for a Directive on adapting non contractual civil liability rules to artificial intelligence

Unfortunately there is no video recording available for this LTS.

Iris Wuisman, Professor Company  Law

Europe is a front-runner in its attempts to regulate AI-systems. In 2021 the European Commission presented a proposal for the European AI Act. The rules follow a risk-based approach and establish obligations for providers and users depending on the level of risk the AI can generate. The proposal has been discussed intensively by many stakeholders and big steps are taken recently in the legislative process with the adoption of a general approach by the Council in December 2022 and adoption of a draft negotiating mandate by the Internal Market Committee and the Civil Liberties Committee of the European Parliament very recently on 11 May 2023. This mandate needs to be endorsed by the whole Parliament, with the vote expected during the 12-15 June session.

In this SAILS Lunch seminar Prof. dr. Iris Wuisman will discuss the key aspects of the proposal and the compromise texts of the Council and the adopted text by the responsible EP committees (LIBE & IMCO) in which many amendments to the Commission’s proposal are included. 

Ksenia Fedorova, University Lecturer at Humanities

The talk addressed the intersection between epistemological, aesthetic and cultural challenges presented by AI. Machine learning became not only an instrument but an object of acute critical reflection for art. The questions at stake include the ethical and political dimensions of the data sets used to train AI models. Art offers ways to deconstruct the premises behind the AI technology and to discover its creative potential and limitations. Critical artistic practices, such as by Lauren McCarthy, Adam Harvey, Trevor Paglen, Helena Nikonole, go far beyond the dominant forms of AI-driven aesthetics, inviting to demystify the technological processes, counter the problematic (techno-solutionist) narratives and resist new forms of cultural commodification.

Computational capabilities push us to reconsider human cognitive representations, the role of contingency and distinctions between the human and the nonhuman types of intelligence. Fedorova thus discussed another thought-stimulating dimension of critical art with and about AI, namely the psychological effects that emerge due to human projections and expectations in interaction with machines.

Francien Dechesne, Associate Professor at the Faculty of Law

While the SAILS network may be initially perceived as a vehicle for research around AI, it soon became clear that there is unique potential for interfaculty and interdisciplinary synergies in education within Leiden University. For the current academic year, this potential has successfully taken shape in the minor program “AI and Society” which productively brought together many of the parties active in SAILS. Now the adventure of the first run of the program has been completed, Francien Dechesne used the SAILS lunch seminar to present the curriculum, share some of experiences of teachers as well as reflections from the students who participated, and look ahead to the further development of the program and wider unique education opportunities around AI for Leiden.

Max van Duijn, Assistant Professor at LIACS

The recent “generative turn” in Artificial Intelligence has struck many of us, both inside and outside the directly involved research fields. In the realm of natural language generation, OpenAI’s ChatGPT has demonstrated striking abilities to engage in convincing conversations on virtually any topic. Over the past months, researchers have debated extensively to what degree this model “understands” things about the world, including the social-cognitive realm of beliefs, desires, intentions, etc. We conducted various experiments using standardised “Theory of Mind” tasks with GTP3 and, in parallel, with children aged 6-9. Max van Duijn presented some of the results an discuss what he thinks they mean, while placing them in the broader context of current discussions about emergent cognitive capacities in generative language models.

Olga Gadyatskaya, Assistant Professor at LIACS.

Mobile devices have become indispensable in the last decade, and the world has seen an explosive growth of mobile application markets. This growth has attracted various adversaries, and nowadays the threat of mobile malware has become one of the key cyber security challenges. Given the deluge of mobile apps to check, thorough manual inspection is no longer feasible. Therefore, machine learning and AI-based approaches have emerged as the solution. Still, it is not satisfactory to only be able to classify an app as malicious or benign: security analysts need to understand why this classification decision was taken. Explainable AI (XAI) methods are then applied for interpreting the classification model and advising the security analysts why a certain sample is recognized as malicious.

In this talk, Olga introduced the Android malware detection problem and the existing XAI approaches applied in this area. She then discussed domain-specific requirements for XAI methods that security analysts have and formulated properties that express these requirements. She also compared several established XAI methods based on these desired properties, and sketched future XAI research directions that can help our society to combat malware. 

In this seminar, Joost Broekens and the participants discussed interactive GPTs (chatGPT) and their impact on the study of intelligence and AI, a series of prompt-based case studies. 

Marius Staring, professor Machine Learning for Medical Imaging at the Department of Radiology, Leiden University Medical Center (LUMC).'

The AI4MRI ICAI lab: MRI made faster, cheaper and better

'We, a collaboration between LUMC, Leiden University and Philips, recently received funding to start a new lab with the aim to accelerate and improve magnetic resonance imaging (MRI) with the help of AI. (MRI) is a widespread medical imaging modality, and a key instrument for disease diagnosis, prognosis and follow-up. MRI is however also slow to acquire, susceptible to patient movement in the scanner, and considered uncomfortable and stressful by patients.

In this talk I will gently go over the main MRI and AI principles that we leverage to accelerate the imaging, and introduce the lab and lab goals.'

Ad Maas, Curator of Modern Sciences in Rijksmuseum Boerhaave and Professor of Museological Aspects of the Natural Sciences at Leiden University.

'In this presentation I will give a look behind the scenes of brAInpower, an exhibition on Artificial Intelligence in Rijksmuseum Boerhaave, the Dutch national museum of the history of science and medicin in Leiden. At the first sight, an exhibition on this topic seems quite challenging. After all, algorithms are invisible. What kind of telling objects can be displayed?  I will desrcibe how we dealt with this challenge, give an overview of the themes and key-messages of the exhibition and set out what we think the visitors will take away from it.'

Mark de Rooij, SAILS Professor of Artificial Intelligence and Data Theory at the  Leiden Institute of Psychology

Abstract: 

Multivariate data are collected in many scientific disciplines. Often the data can be partitioned in a set of response variables and a set of predictor variables. When the response variables are dichotomous, logistic models are a natural choice for the analyses. In this presentation, we show an analysis framework for multivariate binary response variables with or without predictor variables, where we also distinguish between two types of response processes: the dominance and the single-peaked process. Dominance response processes arise when assessing skills or abilities, whereas single-peaked response processes arise in assessing preferences or attitudes. Models are defined in low-dimensional Euclidean spaces, which allow for visualization of the results. We show that estimation of the model parameters can be performed using majorization-minimization algorithms, where the negative log-likelihood is majorized by a least squares function. We illustrate our methods with a few empirical data sets.

Mike Preuss is assistant professor at LIACS, the Computer Science department of Leiden University. He works in AI, namely game AI, natural computing, and social media computing. Giulio Barbero is a PhD with Mike Preuss.

Abstract: 

Game environments are among the most popular and interactive test cases for all kinds of new AI algorithms. Not only did they help forge the AI hype we currently experience, they continue to fuel technological improvements that benefit future AI research. Additionally, games are increasingly used as research tools to engage participants or to collect data on a wider scale. On a commercial level, games continue to attract larger audiences that rival other forms of entertainment. The experiences they provide are becoming more diverse, involving trends such as increased open-world environments, hybrid gaming, and procedural content generation. In this talk, Mike Preuss and Gulio Barbero present the Game Research Lab of Leiden University and its impact on research and education.

Bart Barendregt, Professor by special appointment Anthropology of Digital Diversity / Scientific director

Bart Barendregt is scientific director of the Institute of Cultural Anthropology and Development Sociology (CADS) and a professor and UNESCO Chair of Digital Diversity.  He is currently PI of the NWO-VICI project One Among Zeroes |0100| Towards an Anthropology of Everyday AI in Islam. 

Digital technologies and religion have more in common than meets the eye, as they both produce futures in both professional and popular imagination. Secularists have long described religion scathingly, as ‘backward’, and the fate of faith as sealed. If ever absent, faith is now back in the public arena, and its future is digital. Islam offers a thought-provoking case, not only as major world religion, but also as the ‘liberation theology’ of the global South. The ideal of an Islamic information society spearheading the latest technologies offers a model for an alternative to the dominant pathways of digital transition: Californian ‘big tech’, the Chinese social credit system, or European regulation.  

The |0100| Team investigates the Islamic Information Society’s most controversial component – artificial intelligence – with multimodal and mixed methods, comparing and contrasting narratives and imagery of digital religious futures in the national settings of Malaysia, Indonesia and Singapore, each with a considerable yet differently positioned Muslim population. It situates ethical dilemmas surrounding algorithms, bots and deep learning by ethnographically and longitudinally observing and interviewing makers and users. Innovatively using a historical analysis of future-making discourse, we probe big and ‘thick data’ (‘in situ’ and through digital ethnography) and use infographics, animations and comics to both map and represent ‘scripted futures’. |0100| delivers salient ethnographic portraits as well as an overarching socio-anthropological analysis, of what digital everyday religion is like, how Southeast Asians use AI ‘in the wild’, and how digital technology contributes to exciting societal experiments and ethical dilemmas. 

Studying religious digital futures anthropologically means openly and seriously contemplating the multiple possible directions of digitalizing societies. It raises public awareness of the moral decisions implied in how and why we let technology shape our lives, and reminds experts and policymakers about the fact that the future is always configured here-and-now. 

Serena Viti, Professor of Molecular Astrophysics

Astrochemistry is the study of molecules in space. Molecules are present in the dense gas in all galaxies and, beside having a key role in the formation and shaping of such galaxies, they are also ideal tracers of their physical characteristics. In order to draw the real potential of molecules as tracers of the dense gas in  galaxies, accurate estimates of their abundances  as a function of all the physical parameters that influence their chemistry must be obtained.  However understanding the physical conditions in the molecular gas in galaxies is often an inverse problem subject to complicated chemistry that varies non-linearly with both time and the physical environment.  In this talk I will present examples of how we can use statistical and AI techniques to interpret astrochemical observations as well as improve astrochemical modeling. 

Carolina Mattsson, Postdoc with the Computational Network Science group within LIACS.

The financial connections underlying modern economies are often largely unseen; data on company-company trade is rarely collected, and mostly proprietary. Here we situate so-called production networks within a wider space of networks that are different in nature, but similar in local connectivity structure. Through this lens, we study a regional and a national network of inferred trade relationships reconstructed from Dutch national economic statistics and re-interpret prior empirical findings. We find that company-level production networks have “functional” local connectivity structure, as previously identified in protein-protein interaction (PPI) networks. Functional networks are distinctive in their over-representation of closed squares, which we quantify using an existing measure called spectral bipartivity. Shared local connectivity structure lets us ferry insights between domains: PPI networks are shaped by complementarity, rather than homophily, meaning that companies are especially similar to their close competitors, not to their trading partners. Our findings have practical implications for downstream machine learning tasks involving financial connections among companies, such as link prediction and network embeddings. Moreover, production networks are integral to economic dynamics and our findings are of particular relevance to industrial climate policy.

Daan Pelt, Assistant professor in Data Science, Liacs

State-of-the-art deep learning techniques such as convolutional neural networks tend to require large amounts of training data to be able to perform the required task accurately. In scientific imaging problems, this training data is often obtained through manual annotation of relevant data by a human domain expert. This process typically requires expert domain knowledge, and is often time-consuming, tedious, subjective, and prone to errors. These challenges can lead to inaccurate trained networks and high costs for producing training data. Therefore, a current active field of research is to develop techniques to (1) limit the amount of required manual annotation, (2) improve the reliability of manual annotations, and (3) reducing the impact of inconsistencies in the annotations. In this talk, I will highlight a few recent results for this goal of optimizing manual annotation for producing training data.

Matt Young, Assistant Professor of Public Administration

This project investigates the consequences of implementation choices when using machine learning (ML) to automate public sector decision making. ML has proven to be extremely susceptible to artifacts in training data that introduce bias and lead to suboptimal decision output. Public sector organizations have a responsibility to avoid making biased decisions, both because of their mandate to protect individual rights, and because of the power that they wield over people’s lives. Previous work on representative bureaucracy finds that administrators are more likely to treat individuals fairly when they share characteristics with the population they serve. Thus, all else equal, administrative data generated under conditions of active representation should contain less embedded bias, improving ML performance when used as training data. We propose a 3x3 treatment-control design to test different ML architectures trained on administrative data previously employed to identify and estimate the effect of active representation in the context of education policy on future automated decision-making.

Boudewijn Lelieveldt, Professor Biomedical Imaging

This presentation discusses novel visual analytics and non-linear dimensionality reduction techniques for large, high dimensional datasets. Focusing on the non-linear embedding technique tSNE, we developed Dual tSNE and linked-view tSNE to enable fast and interactive identification of clusters and functionally interesting feature sets. Moreover, we developed spatially mapped tSNE that integrates spatial image information in the tSNE map analysis. Finally, we developed Hierarchical Stochastic Neighbor Embedding, that scales to millions of data points while preserving the manifold structure of the full dataset. Applications of these techniques will be highlighted in two application examples: 1) HDPS: a generic plugin system for fast and interactive analysis of large high-dimensional data sets, 2) Data viewers developed for interactive exploration of the single-cell data resources of the Allen Institute for Brain Science in Seattle.

Lotte van Dillen, Associate Professor at Leiden University

Footage from cameras that have recorded an arrest, chase, confession or crime scene in real time is also called 'evidence verité'. Many people think that these kinds of images show what really happened; "camera images do not lie". This ties in with naïve realism: the human tendency to believe that images show reality. Thanks to this naïve realism, we often find it unnecessary to discuss the meanings of camera images or a photo: the images speak for themselves. But images don't speak for themselves. The viewers do this, they interpret the image. What an image shows (and what not), how it is presented, and in which context, all affect the viewer's interpretation. The question that arises is when this interpretation is useful, and when it is harmful for the process of fact finding. In this presentation Lotte discussed insights from two (ongoing) collaborative, mixed-methods research projects for Police and Science, in which we examined this question and explore how the use of camera evidence in investigative policing can be more effective and less biased. Questions that Lotte hoped to discuss with the audience are whether, and when AI can be a helpful means to reduce unwanted bias.

Suzan Verberne, Associate Professor at LIACS

We typically associate machine learning with classification, regression, and clustering. But some machine learning tasks are extraction tasks: we have a sequence of data and we need to extract the relevant information from it. In text data, the relevant information we are looking for are typically entities and relations between them: names of people and events in news texts, proteins and genes in biomedical data, or artefacts and locations in archaeological reports (as we saw in the seminar talk by Alex Brandsen). In this talk I will introduce information extraction and the common machine learning approach to information extraction from text. I will briefly discuss three different projects addressing information extraction in different domains, among which the PhD project of Anne Dirkson in which we have developed text mining techniques to process and extract information from the large volume of messages on a patient forum. Specifically, we have mined side effects of medications, and the coping strategies of patients who suffer from these side effects. I will show the challenges and results of this project.

Suzan Verberne is an associate professor at the Leiden Institute for Advanced Computer Science of Leiden University. She is group leader of Text Mining and Retrieval in which she supervises 7 PhD students. She obtained her PhD in 2010 on the topic of Question Answering systems and has since then been working on the edge between Natural Language Processing and Information Retrieval. She has been involved in projects involving a large number of application domains and collaborations: from art history to law and from archeologists to medical specialists. Her recent work centers around interactive information access for specific domains. She is also involved in a number of projects on social media mining.

Peter Verhaar, Digital Scholarship Librarian and University Lecturer at the Leiden University Centre for the Arts in Society.

As is the case in virtually all academic disciplines, humanities scholars are increasingly trying to harness the manifold possibilities associated with AI. The emergence of tools and algorithms based on machine learning and deep learning have pushed researchers to experiment with data-rich approaches which can help to expose properties of cultural and historical objects they could never observe before, moving beyond the ‘human bandwith’. The transition from mere data creation to actual analysis continues to pose challenges, however. In this presentation I want to discuss two central caveats that need to be taken into account by humanities scholars who aim to work with methods based on AI, and who aim to integrate the outcomes of these methods into their research.

A first important challenge can be created by a lack of explainability of such results. Existing AI algorithms tend to focus first and foremost on the development of models for the classification of specific objects, and the logic underlying such models often receives much less attention. The type of learning that is implemented within deep learning algorithms also differs quite fundamentally from the ways in which humanities scholars have produced knowledge traditionally. During recent years, a number of techniques have been developed, fortunately, to clarify the steps that are followed by algorithms during the generation of predictions and classifications. Such techniques to enhance the explainability of AI algorithms can ultimately help to reconcile methodologies based on AI with the more conventional forms of humanistic thinking.

A second challenge results from the fact that the data sets that are used as training data are often biased. Whereas humanities scholars typically aim to contextualise and to explain events, objects and phenomena by considering these from many different perspectives, the ‘ground truth’ that is used to train models usually reflects one perspective only. It is clear that such biased datasets can have important ramifications for marginalised communities and that they may reinscribe existing social and political inequalities.

Jan Willem Erisman, Professor Environmental Sustainability at Universiteit Leiden

The Liveable Planet programme is one of the eight interdisciplinary programmes that were launched at Leiden University in 2020, SAILS being one of the others. Leiden’s Liveable Planet programme aims to combine scientific, policy, socio-cultural and historical/archaeological research at Leiden University into coherent research with which we can tackle the major challenges of a transition to a habitable planet with ecological sustainability. The programme will  serve as a hub for the wide range of relevant research carried out within Leiden University and welcomes interaction with colleagues interested in contributing to the initiative within as well as outside of Leiden University.

The Netherlands is in the top 5 of most happy people. At the same time, we are experiencing several crises such as the nitrogen, climate, biodiversity, housing, and sustainable energy crisis. The current policy mainly addresses short-term problems in isolation and does not look far ahead, with new problems looming on the horizon. With the Liveable Planet program we will stimulate community based approaches in the global context to help address these crisis. We use the Sustainable Development Goals for 2030 as a starting point and will contribute to achieving these goals at all scales. This requires multidisciplinary  approaches, new methods, instruments and big data. In this lunch presentation I will give an overview of the Liveable Planet programme and provide an overview of the challenges and opportunities where AI might play a significant role.

Alex Brandsen, Postdoc researcher in Digital Archaeology at the Faculty of Archaeology.

The amount of archaeological literature is growing rapidly. Until recently, these data were only accessible through metadata search. We implemented a text retrieval engine for a large archaeological text collection (~658 Million words). In archaeological IR, domain-specific entities such as locations, time periods, and artefacts, play a central role. This motivated the development of a named entity recognition (NER) model to annotate the full collection with archaeological named entities.
In this talk, we present ArcheoBERTje, a BERT model pre-trained on Dutch archaeological texts. We compare the model's quality and output on a Named Entity Recognition task to a generic multilingual model and a generic Dutch model. We also investigate ensemble methods for combining multiple BERT models, and combining the best BERT model with a domain thesaurus using Conditional Random Fields (CRF).
We find that ArcheoBERTje outperforms both the multilingual and Dutch model significantly with a smaller standard deviation between runs, reaching an average F1 score of 0.735. 
Our results indicate that for a highly specific text domain such as archaeology, further pre-training on domain-specific data increases the model’s quality on NER by a much larger margin than shown for other domains in the literature, and that domain-specific pre-training makes the addition of domain knowledge from a thesaurus unnecessary. At the end of the presentation, a short demonstration of the entity search system is given. 

Peter van der Putten, Assistant Professor AI at LIACS

Abstract: 

What is it what makes us uniquely human? Is it intelligence, or something else? In this talk I will give a broad overview of my research theme and practice Artificial X: investigating human qualities such as intelligence, but also creativity, emotions, curiosity, bonding, obedience or even topics such as morality and religion, through an artificial creature lens. I will illustrate this with a kaleidoscopic sampling of projects from previous years, ranging from research to creative student works, as well as a personal project currently on display at Museum De Lakenhal and ZKM Karlsruhe. These projects help us reflect on what can we learn from these bots about ourselves and what not, encourage general public debate and speculate on what our joint future with Artificial X may look like.   

André Krom - LUMC

Abstract: 
In this talk I will present an overview of common and pressing ethical issues and dilemmas faced by researchers working on potential AI applications for health care purposes. This will be the main part of my contribution to the webinar. Knowledge of such ethical issues and dilemmas, however, is one thing. Having options for action to actuallt deal with them, is another, and equally important. In the interest of providing options for action, I will therefore briefly introduce the key features of a constructive approach in applied ethics called "guidance ethics". While this approach is typically used to structure conversations about ethical questions concerning the application of AI-systems, I will argue and show that it is helpful in the context of facing ethical issues and dilemmas in AI research as well.

Wouter Verschoof - Van der Vaart, Post-doc researcher at the Faculty of Archaeology

Abstract:

The manual analysis of remotely-sensed data, i.e., information about the earth obtained by terrestrial, aerial, and spaceborne sensors, is a widespread practice in local and regional scale archaeological research, as well as heritage management. However, the ever-growing set of largely digitally and freely available remotely-sensed data, creates new challenges to effectively and efficiently analyze these data and find and document the seemingly overwhelming number of potential archaeological objects.  

Therefore, computer-aided methods for the automated detection of archaeological objects are needed. Recent applications in archaeology mainly involve the use of Deep Learning Convolutional Neural Networks (CNNs). These algorithms have proven successful in the detection of a wide range of archaeological objects, including prehistoric burial mounds and medieval roads. However, the use of these methods is not without challenges. Furthermore, in archaeology these approaches are generally tested in an (ideal) experimental setting, but have not been applied in different contexts or 'in the wild', i.e., incorporated in archaeological practice. Even though the latter is important to investigate the true potential of these automated approaches.  

In this talk we will explore some of the opportunities and limitations of using CNN-based object detection in archaeological prospection and the potential—on both a quantitative and qualitative level—of these methods for landscape or spatial archaeology. 

Towards reliable and trustworthy AI systems

Jan van Rijn, assistant professor in Artificial Intelligence.

Abstract: 

The enormous potential of artificial intelligence is like a two-sided sword for society. When applied correctly, these systems can make a positive difference in our daily lives. On the other hand, sloppy deployment can lead to severe damage and even dangerous situations. This consideration becomes more important as artificial intelligence systems get further integrated into our society. As such, there is an obligation for the research community to develop methods that ensure safe deployment and verification techniques, to ensure beneficial applications.

Machine learning models (in particular deep neural networks) are known to be vulnerable to adversarial attacks. By injecting Gaussian noise into the input, the model can be influenced to make a pre-determined decision. The research field of neural network verification develops techniques that determine for a given model how vulnerable it is to such adversarial input perturbation. However, these techniques require a lot of domain expertise and are computationally expensive.

In this talk, I will present our recent work on neural network verification, that can determine whether such a network is vulnerable or safe. I will overview the state of the art neural network verifiers, and explain what their strengths and weaknesses are. These can be sped up by applying advances in hyperparameter configuration. When selecting the right hyperparameter configuration for such a validator, the validation process can be sped-up drastically, and validation resources can be utilized more focused, leading to more reliable AI systems.

What to Include in a Future Children's Book about AI

Bas Haring, professor of Public understanding of Science at LIACS.

Abstract: 

Although some might know me as a philosopher, I actually have a background in artificial intelligence. After my Ph.D. I didn't do much with that field, or so it seems. I wrote popular scientific books about e.g. evolution, biodiversity and even about economics. However, all these topics are in fact linked to artificial intelligence — I realised later in my career. In this meeting I will tell you how these topics are linked.

Rineke Verbrugge, Professor Logic and Cognition at Groningen University

Abstract:

When engaging in social interaction, people rely on their ability to reason about other people’s mental states, including goals, intentions, and beliefs. This theory of mind ability allows them to more easily understand, predict, and even manipulate the behavior of others. People can also use their theory of mind to reason about the theory of mind of others, which allows them to understand sentences like “Alice believes that Bob does not know that she wrote a novel under pseudonym”. But while the usefulness of higher orders of theory of mind is apparent in many social interactions, empirical evidence so far suggested that people often do not use this ability spontaneously when playing games, even when doing so would be highly beneficial. 

In this lecture, we discuss some experiments in which we have attempted to encourage participants to engage in higher-order theory of mind reasoning by letting them play games against computational agents: the one-shot competitive Mod game; the turn-taking game Marble Drop; and the negotiation game Colored Trails. It turns out that we can entice people to use second-order theory of mind in Marble Drop and Colored Trails, and in the Mod game even third-order theory of mind.

We discuss different methods of estimating participants’ reasoning strategies in these games, some of them based only on their moves in a series of games, others based on reaction times or eye movements. In the coming era of hybrid intelligence, in which teams consist of humans, robots and software agents, it will be beneficial if the computational members of the team can diagnose the theory of mind levels of their human colleagues and even help them improve.

James W. McAllister, Professor in History and Philosophy of Science, Leiden University

Abstract: 
Empirical data -- the results of observations and measurements -- contain information about the world. But how much information do they contain, and what does this tell us about the structure of the world? A traditional reply to these questions is that an empirical data set contains a single pattern, and that this shows that the world has a unique structure. This reply suggests that scientific laws and theories constitute, in the terms of algorithmic information theory, a lossless compression of empirical data. I argue to the contrary that scientific practice depends on decomposing empirical data sets into two additive components: a simple pattern, which corresponds to a structure in the world, and a residual noise term. This view leads to two intriguing implications. First, if the noise term is algorithmically incompressible, then empirical data sets as a whole are also incompressible. Second, since it is possible to decompose a data set into a pattern and noise in infinitely many different ways, and since each of these decompositions has equal claim to validity, then empirical data are evidence that the world contains an infinite amount of structure

Arko Ghosh, Assistant Professor at the Faculty of Social and Behavioural Sciences.

Abstract:

The time-series of smartphone touchscreen interactions (tappigraphy) may help resolve the systematic links between brain functions and behavior in the real world. In this talk, I will provide an overview of tappigraphy, and how we are applying it to unravel human behavior in health and neurological disease. My talk will involve life span measurements, cognitive tests, brain implant recordings, and long-term behavioral monitoring. I will present straightforward statistical models linking tappigraphy to brain functions in healthy people, and machine learning approaches that help infer brain status based on tappigraphy inputs in Epilepsy and Stroke. These studies vividly demonstrate the potential of tappigraphy to investigate fundamental neuro-behavioral processes relevant to the real world.

Linnet Taylor, Professor of International Data Governance at the Tilburg Institute for Law, Technology, and Society (TILT)

Abstract:

The process of setting rules and norms for computing processes and applications has been dominated by requirements engineering and formalisable, rather than ‘thick’ interpretations of central concepts including fairness, responsibility, trust and participation. Yet computing science experts and other disciplines such as law and philosophy often understand these terms very differently. These differences in understanding can create productive friction and discussion amongst experts with very different backgrounds and orientations, but can also constitute gaps that lead to governance-by-default, where instead of creating architectures for the control and shaping of digital power and intervention, disagreement on fundamental concepts delays action. This talk will explore whether these diverging understandings represent fundamental incompatibilities between disciplinary worldviews, what the effects of the resulting faultlines are in terms of thes target and aims of governing data and AI, how we can recognise productive disjunctures. I will look particularly at the current politics of AI, and ask whether there are ways to govern technology when different groups are locked in opposition around core concepts and assumptions which each consider non-negotiable.

Gerhard de Kok, Teacher at the Institute of History, Leiden University

Abstract:

Handwritten Text Recognition (HTR) is revolutionizing historical research. Models trained using neural networks can now read seventeenth and eighteenth century handwriting better than most humans alive today. Such models have already automatically transcribed vast archives, including those of the Dutch East and West India Companies (WIC and VOC). The resulting transcriptions are not perfect, but they open up new avenues for research. For the first time in history, it is possible to search through these archives with full text search. The millions of pages of transcriptions also invite further exploration through the use of NLP techniques. A first experiment with word embeddings has already led to some promising results.

Karën Fort, Associate Professor at Sorbonne Université, France

Abstract: 

In the last decade, AI, and especially Natural language processing (NLP) applications have become part of our daily lives. Firms like Google, Facebook or Amazon have devoted huge efforts into research and are present in all our conferences. The distance between researchers and users has shrunk and a number of ethical issues started to show: stereotypes are repeated and amplified by machine learning systems, AI is used for ethically-questionable purposes like automatic sentencing, and more or less experimental tools are forced on users without taking their limitations into account. In this presentation, I'll detail some of the issues we are faced with, and I'll propose a  systemic view on the subject that allows to uncover some blind spots that have to be discussed.

Rob Saunders, Associate Professor at LIACS, Leiden University

Abstract:

Creativity is one of the most highly prized faculties of human intelligence. The history of speculation about intelligent machines is mirrored by a fascination with the possibility of mechanical creativity. From the myths and legends of antiquity to the Golden Age of Automata in the 18th Century, the achievements of mechanical wonders were often paired with amazement at the performance of apparently creative acts. During the 20th Century the fascination with creative machines continued and at the dawn of the Computer Age the prospect of computationally modelling creative thinking was proposed as one of the “grand challenges” in the prospectus for the field of Artificial Intelligence. In the past 60 years, the field of Artificial Intelligence has seen significant progress in realising the goal of building computational creativity, from early Discovery Systems to the latest advances in Deep Learning. Like intelligence, however, the notion of creativity is an essentially social construct. Much work remains if creative machines are ever to become a reality, both in terms of technical advances and the integration of such machines into society. In addition, the development of machines capable of acting in ways that would be considered creative if performed by a human, will challenge our understanding of what it means to be creative. This talk will explore the history of creative machines and the prospects for the future of computational creativity research.

Marco Spruit, Professor Advanced Data Science in Population Health, Leiden University

Abstract:

In this overview talk, I will first position the research domain of Translational Data Science, in the context of the COVIDA research programme on Dutch NLP for healthcare. Then, I will present our prognostic study on inpatient violence risk assessment by applying natural language processing techniques to clinical notes in patients’ electronic health records (Menger et al, 2019). Finally, I will discuss followup work where we try to better understand the performance of the best performing RNN model using LDA as a text representation method among others, which reminds us once more of the lingering issue of data quality in EHRs.

Marjolein Fokkema, Assistant Professor at Psychology, FSW, Leiden University

Abstract:

Prediction Rule Ensembling (PRE) is a statistical learning method that aims to balance predictive accuracy and interpretability. It inherits high predictive accuracy from decision tree ensembles (e.g., random forests, boosted tree ensembles) and high interpretability from sparse regression methods and single decision trees. In this presentation, I will introduce PRE methodology, starting from the algorithm originally proposed by Friedman and Popescu (2008). I will show several real-data applications, for example on the prediction of academic achievement and chronic depression. I will discuss several useful extensions of the original algorithm which are already implemented in R package ‘pre’, like the inclusion of a-priori knowledge, unbiased rule derivation, and (non-)negativity constraints. Finally, I will discuss current work in which we leverage the predictive power of black-box models (e.g., Bayesian additive regression trees, deep learning) to further improve accuracy and interpretability of PRE.

Johannes Schmidt-Hieber, Mathematical Institute, Leiden University

Abstract:

Recently a lot of progress has been made regarding the theoretical understanding of machine learning methods. One of the very promising directions is the statistical approach, which interprets machine learning as a collection of statistical methods and builds on existing techniques in mathematical statistics to derive theoretical error bounds and to understand phenomena such as overparametrization. The talk surveys this field and describes future challenges.

Mitra Baratchi, Assistant Professor at LIACS, Leiden University

Abstract:

Spatio-temporal datasets (e.g., GPS trajectories, Earth observations) are ubiquitous. Algorithms for effective and automated processing of such data are relevant from various applications, from crowd movement analysis to environmental modelling. These algorithms need to be designed considering the fundamental aspects of the underlying spatio-temporal processes (e.g., the existence of spatial and temporal correlations) and be robust against various ubiquitous data imperfection issues. In this talk, I will introduce the field of spatio-temporal data mining and talk about crucial open research challenges for making use of such data.

I would also like to discuss the vision of creating a “data observatory” to address various important research challenges in multi-disciplinary research. The data observatory aims to bring together datasets (the observations), AI algorithms (the tools), and expertise (the humans) in a well-equipped setting that facilitates a collaborative investigation.

Gerard van Westen, Professor of Artificial Intelligence & Medicinal Chemistry, LACDR, Leiden University

Abstract:

Drug discovery is changing, the influence and catalytic effect of Artificial Intelligence (AI) cannot be denied. History dictates this new development will likely be a synergistic addition to drug discovery rather than a revolutionary replacement of existing methods (like the history of HTS or combichem, a new tool in the toolbox). As more and more scientific data is becoming public and more and more computing power becomes available the application of AI in drug discovery offers exciting new opportunities.

Central to drug discovery in the public domain is the ChEMBL database which provides literature obtained bioactivity data for a large group of (protein) targets and chemical structures.[1, 2] Machine learning can leverage this data to obtain predictive models able to predict  the  activity probability of untestedchemical structures  contained within the large collections of chemical vendorson the basis of the chemical similarity principle. [3, 4]

In this talk I will give an overview of research going on at the computational drug discovery group in Leiden. Central in our research is the usage of machine. I will highlight some examples we have published previously and finish with an outlook of cool new possibilities just around the corner.[5, 6]

References
1. Sun, J., et al., J. Cheminf., 2017. 9, 10.1186/s13321-017-0203-5
2. Gaulton, A., et al., Nucleic Acids Res., 2012. 40, 10.1093/nar/gkr777
3. Bender, A. and R.C. Glen, Org. Biomol. Chem., 2004. 2, 10.1039/b409813g
4. Van Westen, G.J.P., et al., Med. Chem. Commun., 2011. 2, doi:10.1039/C0MD00165A
5. Liu, X., et al., J. Cheminf., 2019. 11, 10.1186/s13321-019-0355-6
6. Lenselink, E.B., et al., J. Cheminf., 2017. 9, 10.1186/s13321-017-0232-0

Nele Mentens, Professor of Computer Science, LIACS, Leiden University

Abstract: 

Artificial Intelligence plays an important role in the protection of electronic devices and networks. Examples of domains in which AI has shown to lead to better products and protection mechanisms, are the security evaluation of embedded and mobile devices, and the detection of attacks in IoT and IT networks. Besides the added value that AI brings, there are also a number of pitfalls with respect to the privacy of users whose personal data are processed, and the confidentiality of the models that are employed. This talk will give an overview of these opportunities and challenges. 

Bart Custers, Professor of Law and Data Science, Leiden University

Abstract:

AI is developed and used for many good causes, but increasingly criminals also make use of developments in AI. In this presentation, examples of crime are examined that involve AI and related technologies, including big data analytics, A/B optimization and deepfake technology. Typically such technologies can enhance the effectiveness of crimes like ransomware, phishing and fraud. Next, it is discussed how AI related technologies can be used by law enforcement for discovering previously unknown patterns in crime and empirical research on what works in sanctioning. Examples of novel patterns are presented as well as existing sophisticated risk assessment systems. From a procedural criminal law perspective, i.e., when investigating crime, AI technologies can also be used both in providing cues during criminal investigations and in finding evidence. Approaches in predictive policing are investigated as well as the potential role of existing cyber agent technology. With regard to finding evidence, advanced data analytics can prove to be helpful for finding the proverbial needle in the haystack, providing Bayesian probabilities and building narratives for alternative scenarios. For all these developments, legal issues are identified that may require further debate and academic research.

Dr Oleh Dzyubachyk, Post-Doctoral Researcher at LUMC, Leiden University

Abstract: 

Electron microscopy (EM) is an imaging modality that has vast potential for becoming one of the primary beneficiaries of the advance of machine learning. In my talk I will first introduce to you this imaging modality and provide a few examples of data quantification needs. Next, I will describe our recent developments that enabled applying machine learning methodology to our in-house data and preliminary results of the mitochondria quantification project. Finally, I will share with you my ideas about potential directions for future research.

Daan Pelt, Assistant Professor at LIACS, Leiden University

Abstract: 

In recent years, convolutional neural networks (CNNs) have proved successful in many imaging problems in a wide variety of application fields. However, it can be difficult to train and apply existing CNNs to scientific images, due to computational, mathematical, and practical issues. In this talk, I will discuss newly developed methods that are specifically designed for scientific images. These methods can accurately train with large image sizes and a limited amount of training data (or no training data at all), and can automatically adapt to various tasks without extensive hyperparameter tuning. The talk will include comparisons between the new methods and existing CNNs, some recent results for real-world scientific problems, and ideas for future work.

Anne Meuwese, Professor Public Law, Leiden University

Abstract: 

The intersection of 'Public Law', 'Governance' and 'Artificial Intelligence' is not limited to the question of how AI can be regulated. AI also has the potential to change certain fundamental processes of the state. This presentation looks at one such process: lawmaking. In what ways could IA change both the process and the outcome of lawmaking by legislators? Among the possible applications of AI in lawmaking discussed are 1) the use of AI in monitoring the effects of legislation ‘ex post’, in particular the potential of AI in identifying regulatory failures, 2) the possible changes in the types of norms used in legislation in sectors in which AI is used to support administrative decision-making (rules vs standards, level of abstraction, ‘granularity’ of norms), 3) the implications of AI for the idea of ‘technology neutrality’ in legislative drafting, 4) the expected increased frequency of legislative projects for which an (AI) system will need to be designed in parallel. To what extent to we see these applications emerging and what are the implications for the fields of public law and governance?

Tuna Kalayci, Assistant Professor at the department of Archaeological Sciences, Leiden University

Abstract:

The historical imagery is invaluable in archaeological research. At the very least, an old photograph offers the first record of an object (ranging from a small pottery piece to a large ancient settlement). In some cases, these images might be the only records left due to destruction of that object. Therefore, it is beneficial for archaeologists to explore these data sources to the full extent. This talk examines one of these sources, CORONA spy-satellite and discusses the results of a CNN model for the automated documentation of ancient settlements in the Middle East. This talk will also include brief evaluations of two potential future projects: Sounds of Leiden (SouL) and Robotics in Archaeology.

Matthijs Westera, Assistant Professor at LUCL, Leiden University

Abstract: 

When we talk, we try to be clear, coherent, relevant, informative and truthful, or at least to appear that way. An audience will expect this, and this expectation constrains their possible interpretations of our utterances. How exactly this works is the topic of the linguistic field of Pragmatics, where a helpful notion has proven to be that of a Question Under Discussion (QUD): a (typically implicit) question to which an utterance must provide some kind of answer. In a coherent discourse, every utterance should address a pertinent QUD, ideally one that was evoked by the preceding discourse. Despite their centrality in the field of Pragmatics, QUDs have received only little attention in Natural Language Processing (NLP), where the vast majority of work on discourse coherence is not QUD-based but relation-based (discourse relations such as 'explanation' and 'elaboration'), and virtually any work on questions concerns, instead, either question answering (given a question, find a suitable answer to it) or 'answer questioning' (given an answer contained in a text, generate a suitable comprehension question for it). I will present my (et al.) ongoing attempts (crowdsourcing and computational modeling) to add QUDs to the NLP toolbox, hoping to receive valuable suggestions for, e.g., possible applications in the various fields represented at SAILS.