OFAI is delighted to announce its 2022 Lecture Series, featuring an eclectic lineup of internal and external speakers.
The talks are intended to familiarize attendees with the latest research developments in AI and related fields, and to forge new connections with those working in other areas.
Most lectures (see prospective schedule below) will take place on Wednesdays at 18:30 Central European (Summer) Time. All lectures will be held online via Zoom; in-person attendance at OFAI is also possible for certain lectures. Talks typically last from 30 to 50 minutes and are followed by up to 30 minutes for questions.
Attendance is open to the public and free of charge. No registration is required.
Subscribe to our newsletter or our RSS feed, or bookmark this web page, to receive connection details for the individual talks.
29 June 2022 @ 18:30 CEST
Scott Patterson (McGill University)
Domesticating Wealth Inequality: Hybrid Discourse Analysis of UN General Assembly Speeches, 1971–2018
The rise of wealth inequality is well-documented, yet robust calls for redistribution have yet to emerge from within the UN system. How can this be? We argue that discourse on wealth inequality has been gradually, but pervasively, "domesticated" in diplomatic venues. We advance this argument through a hybrid discourse analysis of UN General Assembly speeches that adopts complementary machine learning and interpretive techniques. Our study demonstrates the usefulness of hybrid discourse analysis for several tasks in time-series text analysis. First, we use a multiclass Support Vector Machine to discover rhetorical "eras" that emerge in debates over time. Second, we use Concept Mover Distance - a distributed, pre-trained language model - to estimate the extent of engagement with rhetorical tropes that are relevant to wealth inequality. These tropes were selected through an interpretive analysis of exemplary external texts. Third, we cluster states based on patterns of engagement with key tropes, illustrating changes to the rhetorical topography over time. Beyond our substantive contribution, we aim to close the gap between interpretive and computational text analysis and to demonstrate the advantage of placing these approaches on an even footing.
6 July 2022 @ 18:30 CEST
Pamela Breda (University of Applied Arts Vienna)
Feeling for Nonexistent Beings
“The Unexpected” is an artistic research project exploring the impact of Artificial Intelligences (AIs), interfaces and digital assistants, on users’ daily life through theoretical research and practice-based field analysis. The constant growth of AI softwares and virtual assistants employed to enhance wellbeing and overcome mental health issues is creating new dimensions of social and ethical questions about human emotional responses to artificial intelligence. While Microsoft is studying the implementation of AI companions with an emotional connection to satisfy the human need for communication and affection, chatbots and digital avatars are designed in order to support users in overcoming anxiety, loneliness and other stress-related situations. But what are the emotional responses to such interactions? Do we still perceive the AIs as “others” or do we empathise with them as they would with humans? Using an interdisciplinary approach the research will analyze the practical impact of human interactions with digital assistants through the development of a written essay and extensive audio-visual documentation edited in the form of an experimental film.
13 July 2022 @ 18:30 CEST
Brigitte Krenn (OFAI)
Robots as Social Agents: Between Construct and Reality
As soon as we humans encounter other agents, be it our pet animals or robots we collaborate with, we cannot help but act socially and interpret our vis-à-vis as social agents. This is because we are trained as social beings from the beginning of our life. We have learned to interpret nonverbal signals sent by our fellow humans as communicative cues, including facial expressions, gestures and body postures, the direction of eye gaze, as well as proximity relations. Being who we are, we tend to overestimate and misinterpret current robots and AI systems regarding their communicative intents. The talk will address this phenomenon from a perspective of human communication and present examples from a selection of robotics research projects, studying human-robot interaction in different application contexts.
20 July 2022 @ 18:30 CEST
Tristan Miller (OFAI)
What's in a Pun? Assessing the Relationship Between Phonological and Semantic Distance and Perceived Funniness of Punning Jokes
Puns are a form of humorous wordplay based on semantic ambiguity between two phonologically similar words. By using and extending a large annotated corpus of punning jokes, we quantify the phonological and semantic distance between the two words of a pun and assess possible correlations with funniness ratings of the joke. Statistical analyses reveal a significant negative correlation between phonological distance and perceived funniness, which is in line with a longstanding conjecture in humour studies. Interestingly, none of the semantic distance measures we applied showed significant correlations with funniness ratings. We discuss other factors, such as situational context or cultural norms, which may influence the perception of funniness of punning jokes, with a view to guiding future research on this topic.
27 July 2022 @ 18:30 CEST
Katrien Beuls (University of Namur)
Unravelling the Computational Mechanisms Underlying the Emergence of Human-like Communication Systems in Populations of Autonomous Agents
Over the last two decades, important advances in the field of artificial intelligence have led to tremendous progress in many tasks and application domains, including computer vision, robotics and natural language processing. Yet, the communication systems that are used by artificial agents for human-agent and agent-agent communication today are still far removed from exhibiting the expressiveness, flexibility and adaptivity that is found in human languages. This gap may mostly be ascribed to the fact that current communication systems are learned by extracting frequently occurring patterns from huge amounts of annotated data, limiting their applicability to predefined tasks set in stable environments. In this talk, I will present my long-term research programme which takes a radically different approach with the goal of building truly intelligent systems that are capable of adapting to unforeseeable changes in their tasks and environment. Rather than extracting patterns from annotated data, we equip populations of autonomous agents with computational mechanisms that allow them to self-organise an emergent conceptual and linguistic system through communicative interactions. By means of multi-agent experiments, we investigate the mechanisms that are needed for inventing, adopting and aligning transparent languages based on novel compositions of atomic cognitive capabilities that are mastered by the agents. These methodological innovations have the potential to lead to a paradigm shift in the way in which explainable human-agent and agent-agent communication is modelled, both in emergent communication experiments and real-world applications. Such applications include safety assistants (communicating with humans), self-driving vehicles (communicating with each other) and distributed smart devices in a home environment (communicating with humans and each other).
7 September 2022 @ 18:30 CEST
Steffen Eger (Bielefeld University)
Text Generation for the Humanities
In this talk, I will discuss our current work in the context of deep learning based text generation for the humanities. In particular, I will talk about (i) style-conditioned poetry generation, (ii) abstract-to-title generation (considering humorousness as one generation criterion), and (iii) cross-lingual cross-temporal summarization where the goal is to summarize a historical document in another modern language. Time permitting, I will also talk about evaluation metrics for text generation systems.
14 September 2022 @ 18:30 CEST
Antti Arppe (University of Alberta)
Finding Words that Aren't There: Using Word Embeddings to Improve Dictionary Search for Low-resource Languages
Modern machine learning techniques have produced many impressive results in language technology, but these techniques generally require an amount of training data that is many orders of magnitude greater than what exists for low-resource languages in general, and endangered ones in particular. However, dictionary definitions in a comparatively much more well-resourced majority language can provide a link between low-resource languages and machine learning models trained on massive amounts of majority-language data. By leveraging a pre-trained English word embedding to compute sentence embeddings for definitions in a Plains Cree (nêhiyawêwin) dictionary, we have obtained promising results for dictionary search. Not only are the search results in the majority language of the definitions more relevant, but they can be semantically relevant in ways not achievable with classic information retrieval techniques: users can perform successful searches for words that do not occur at all in the dictionary. These techniques are directly applicable to any bilingual dictionary providing translations between a high- and low-resource language.
21 September 2022 @ 18:30 CEST
Roman Pflugfelder (TU München / Technion)
Fragmented Occlusion in Computer Vision
Occlusion is an important and persistent problem in computer vision. Most of the modern, visual recognition algorithms suffer from occlusion. Making algorithms robust to occlusion is challenging, as occlusion is the result of an information loss that emerges from the projection of a three dimensional world onto a two dimensional image. This lecture will introduce fragmented or dynamic occlusion which appears when looking through foliage or when looking through a fence while walking. The problem is known in cognitive science but mostly ignored in computer vision. The current results in cognitive science tell us the importance of the temporal dimension in vision which introduces important clues for visual recognition under fragmented occlusion. Concepts such as the spatiotemporal form integration rely on motion percepts. Inspired by these psychological results, I will introduce a new video processing approach which is named video deocclusion. This approach is able to deocclude a fragmentally occluded, unknown object of interest in a sequence of images. I will also present results of a new algorithm based on deep learning which is able to localise persons behind trees where state-of-the-art object detection algorithms fail. This research is fruitful for a better understanding of the astonishing capabilities of human vision under dynamic occlusion. It is also useful for future applications of computer vision in natural environments.
28 September 2022 @ 18:30 CEST
Raphael Deimel (TU Wien)
Fluent and Intuitive Human–robot Object Handover
Watching humans hand over things, the skill seems trivial. But the smoothness and the ease of interaction is deceiving, as complex, fast and nonverbal communication takes place continuously to negotiate shared information such as exactly when and where to hand over the thing, which handover type to use, who initiates the interaction and who assumes the socially dominant role. All those aspects are influenced by subjective preferences, circumstances and cultural norms. If a robot wants to participate, it too has to perform this interactive, continuos and nonverbal negotiation "dance" competently. It needs to be able to propose courses of action, to acknowledge proposals by affirmative action, to signal the need for more negotiation, to recognize incompatible courses of action and to revert them. Discrete state machines are ill suited to handle the nuances, ambiguity and continuity inherent to communication via body motion, whereas continuous controllers fail to break the problem into easier substeps. Phase-State machines solve this dichotomy: they encode temporal progress, allow for ambiguity and support gradual decision making processes but they also encode arbitrary state graphs.
5 October 2022 @ 18:30 CEST
Christoph Scheepers (University of Glasgow)
The “Crossword Effect” in Free Word Recall: A Retrieval Advantage for Words Encoded in Line with their Spatial Associations
According to the perceptual symbol hypothesis (Barsalou, 1999), word concepts trigger mental re-enactments of perceptual states and actions. While many studies have shown how word concepts modulate sensori-motor responses, it is less well known how sensori-motor actions influence access to word concepts in memory. Here, we investigated how well English words with strong horizontal or vertical associations are retrieved from memory dependent on how they are presented during encoding (i.e., horizontally or vertically printed). Initial pre-testing of 129 candidate words yielded 43 words with a strong horizontal association (e.g., floor, beach, border, etc.) and 51 words with a strong vertical association (e.g., tree, crane, bottle, etc.). These were quasi-randomly compiled into 160 ‘crossword arrays’, each containing 5 horizontally and 5 vertically printed items drawn from the horizontal association word set, as well as 5 horizontally and 5 vertically printed items drawn from the vertical association word set. The main experiment (160 participants) was preregistered on OSF and was introduced to participants as “testing how word arrangements affect subsequent mathematical problem solving”. There were three experimental phases: (1) in the encoding phase, each participant studied a uniquely generated crossword array for ca. 2 minutes; (2) in the following distractor phase, they had to solve simple mathematical equations for 1 minute; (3) in the final (surprize) free recall phase, they were asked to write down as many words as they could remember from the encoding phase. Dependent variables were likelihood of correctly recalled words and retrieval ranks of correctly recalled words in the recall list. Results showed no appreciable effects in retrieval rank, but a clear interaction (p < .001) between word association and word presentation in the likelihood of correct word recall: vertical association words, in particular, were reliably more likely to be recalled correctly when they were presented vertically (i.e., in line with their spatial association) than when they were presented horizontally during encoding. Implications for the perceptual symbol hypothesis will be discussed.
12 October 2022 @ 18:30 CEST
Karën Fort (Sorbonne Université / LORIA)
Ethics and NLP: What we Talk About, What we Don't Talk About Anymore, What we Never Talked About
In recent years, ethics has become a recognized subject in the fields of AI and more particularly in Natural Language Processing (NLP). This recent development is due to several factors, including the fact that NLP has become commercially attractive enough to leave research laboratories and invade our daily lives, with immediately visible consequences for the general public. I will return in this presentation to the evolution of the subject over the last decade, which has seen certain issues become obvious (such as the remuneration of click workers) and no longer be discussed, while others (notably the biases in language models) take center stage, obscuring the most difficult questions.
19 October 2022 @ 18:30 CEST
Benjamin Roth (University of Vienna)
Evaluation and Learning with Structured Test Sets
Behavioural testing – verifying system capabilities by validating human-designed input-output pairs – is an alternative evaluation method of natural language processing systems proposed to address the shortcomings of the standard approach: computing metrics on held-out data. While behavioural tests capture human prior knowledge and insights, there has been little exploration on how to leverage them for model training and development. With this in mind, we explore behaviour-aware learning by examining several fine-tuning schemes using HateCheck, a suite of functional tests for hate speech detection systems. To address potential pitfalls of training on data originally intended for evaluation, we train and evaluate models on different configurations of HateCheck by holding out categories of test cases, which enables us to estimate performance on potentially overlooked system properties. The fine-tuning procedure led to improvements in the classification accuracy of held-out functionalities and identity groups, suggesting that models can potentially generalise to overlooked functionalities. However, performance on held-out functionality classes and i.i.d. hate speech detection data decreased, which indicates that generalisation occurs mostly across functionalities from the same class and that the procedure led to overfitting to the HateCheck data distribution.
25 October 2022 @ 18:30 CEST
Peter Hallman (OFAI)
Comparatives in Arabic
In this talk, I show firstly that English and Syrian Arabic (which is typical of the contemporary Arabic dialects in the relevant respects) share a syntactic constraint on the formation of ‘degree clauses’, the clauses that describe the standard of comparison that in English are introduced by ‘than’, i.e., the bracketed part in ‘Clyde is taller [than Miriam is]’. Secondly, I show that in Syrian Arabic but not English, the relevant restriction also constrains the scope of the comparative itself, effecting the repertoire of possible interpretations for comparative constructions in that language. Consequently, the scope of the comparative and the derivation of the degree clause are syntactically uniform in Arabic but not in English. I offer some speculations on the source of the unexpected non-uniformity of English, which is probably related to differences in the structure of noun phrases between the two languages.
2 November 2022 @ 18:30 CET
Stephanie Gross (OFAI)
Multimodal Human–Robot Interaction in Situated Task Descriptions
Application areas in which robots and humans work together are rapidly growing, for example in private households or in industry. In these areas, a major communication context is situated task descriptions, where humans naturally use verbal as well as non-verbal channels to transmit information to their interlocutor. Therefore to successfully interact with humans, robots need to (1) share representations of concepts with their communication partner, (2) identify human communicative cues and extract and merge information transmitted via different channels, and (3) generate multimodal communicative behavior which is understandable for humans and complies with social norms. In this talk, I will discuss several challenges on the way, including the type of data used for modelling multimodal HRI, generating non-verbal social signals, or multimodal reference resolution in situated tasks.
16 November 2022 @ 19:30 CET
Bernhard Pfahringer (University of Waikato)
Everything Changes, but Your ML Models Stay the Same?
Most Machine Learning assumes i.i.d. data, data that is independently and identically sampled from a distribution. Thus training on a sample and then applying to new samples is a sound procedure. Unfortunately, in the real world, data is almost never i.i.d. This presentation will introduce data stream mining as a way of tackling data streams that change over time. I will in particular highlight opportunities that this scenario offers, that are not present in the static i.i.d. train-then-test setup.
23 November 2022 @ 18:30 CET
Robert Trappl (OFAI)
Postponed until 2023
Date to be announced
Paolo Petta (OFAI)
Title and abstract to be announced.
How to attend: To be announced
