Serge Thill’s research while affiliated with Radboud University and other places

Decision-makers run the risk of relying too much on machine recommendations. Explainable AI, a common strategy for calibrating reliance, has mixed and even negative effects, such as increasing overreliance. To cognitively engage the decision-maker and to facilitate a deliberate decision-making process, we propose a potential `reflection machine' that supports critical reflection about the pending decision, including the machine recommendation. Reflection has been shown to improve critical thinking and reasoning, and thus decision-making. One way to stimulate reflection is to ask relevant questions. To systematically create questions, we present a question taxonomy inspired by Socratic questions and human-centred explainable AI. This taxonomy can contribute to the design of such a `reflection machine' that asks decision-makers questions. Our work is part of the growing research on human-machine collaborations that goes beyond the paradigm of machine recommendations and explanations, and aims to enable greater human oversight as required by the European AI Act.

Machine learning and artificial intelligence currently see rapid developments and growth. In this chapter, we discuss some of the implications of the increasing ubiquitousness of resulting algorithms in technology used to support neurodiverse children in various contexts. The focus is less on discussing specific methods, given the rapid pace at which they currently evolve, and more on an attempt to highlight general principles and concerns. We highlight in particular that although algorithmic approaches can support various stakeholders, such as parents, therapists, teachers, and the children themselves, they cannot and should not replace them. More generally, we will distinguish between different uses of algorithms, namely the description of observable data and the interpretation thereof. We also discuss some of the fundamental challenges, for example in assessing vaguely-defined mental states. Overall, the purpose of this chapter is to create an awareness of the potentials, but also the limitations of machine learning approaches to supporting neurodiverse children in various contexts.

Survival games can be described as video games where the player searches for energy and treasures, while avoiding obstacles and hostile attacks. Ms.Pac-Man and Minecraft are two well-known examples. Currently there are AI models that outperform human players at Ms.Pac-Man, while AI models playing Minecraft above the human level has been a long-standing challenge. This paper concerns what we call pure survival games, which take place in previously unseen worlds containing only energy, water, and obstacles. The challenge of the player is to navigate and survive in those worlds by continuously finding resources and avoiding obstacles. Arguably, animals need to master physical analogues of pure survival games in order to survive and reproduce. Here we begin to explore human and machine performance on pure survival games. We define two games called the Grid game and the Terrain game and two corresponding AI agents based on deep reinforcement learning: the Grid agent and the Terrain agent. We explore to what extent these agents can match human performance and how their performance is affected by variations in their perception, memory, and reward models. We find that (1) the Terrain agent performs above human level, while the Grid agent performs below human level. (2) the smell, touch, and interoception models contribute significantly to the performance of the Grid agent. (3) the memory model contributes significantly to the performance of the Grid agent; and (4) the performance of the Grid agent is relatively stable under three quite different reward signals, including one that rewards survival and nothing else.

Historically it was assumed that computation requires representation, which meant that Post-Cognitivism (PC) naturally rejected computationalism as well. Recent research suggests that computation is possible without explicit representation [7, 6]. This led to the proposal that PC does not necessarily imply anti-computationalism [10]. In this work, we provide initial empirical support that a theory of Comutational Post-Cognitivism (CoPC) [10] is feasible. On the example of Autopoietic Theory (AT) [9] we demonstrate that by choosing a suitable theory of computation such as Mechanistic Computation (MC) [8] as suggested by Villalobos and Dewhurst, it is indeed possible for an observer to interpret a system as being both computationalist and autopoietic. We achieve this using a simulation built on Game of Life (GoL) as virtual environment, as it has previously been shown that autopoietic properties can be identified in GoL [2]. GoL is inspected under the lenses of MC and AT as examples of computationalism and of post-cognitivism, respectively. When we identify something that arguably qualifies as computational mechanism (according to MC) and autopoietic (according to AT), then we have identified something that is qualifies as CoPC.

For a long time it was taken for granted that computation requires representation, which meant that Post-Cognitivism naturally rejected computationalism as well. However, recent advances in neural network research showed that computation is possible without explicit representation. This led to the proposal that Post-Cognitivism does not necessarily imply anti-computationalism. Most of the discussions however remain philosophical in nature. In this work, we provide initial empirical support that such a theory of Comutational Post-Cognitivism is feasible. Specifically, we demonstrate on the example of Autopoietic Theory that, given a suitable theory of computation such as Mechanistic Computation as suggested by Villalobos and Dewhurst, it is indeed possible for an observer to interpret a system as being both computationalist and au- topoietic. We achieve this using a simulation built on Game of Life as virtual environment, as it has previously been shown that autopoietic properties can be identified in GoL.

Artificial intelligence has shown promise for supporting children with neurodevelopmental disorders (NDDs) in educational settings. For such vulnerable population, aspects such as emotion, communication, and motivation are very relevant, but also challenging to be modeled. In this work, we focus on the machine learning technology used in such scenarios, in particular the characteristics of datasets used for model training. We do this by analyzing recent papers on children with NDDs. This will give insight into existing trade-offs, such as data annotation involved in data collection, as well as automation aspects. We also analyze opportunities offered by the functionalities of ML models trained on such datasets. In addition, we point out limitations and future challenges to help advance the area.Keywordsmachine learningdata annotationneurodevelopmental disorderschildreneducationhuman-computer interaction

This work presents a framework for a visual- auditory attention-driven robot eye-head gaze movement, which combines visual and auditory inputs to determine the direction of movement for a social robot. The framework computes the most salient changes in position by considering both visual and auditory cues. The proposed system was implemented on Haru, a tabletop social robot, where eye-head gaze movement was controlled using visual input from a camera positioned above the eyes and auditory input from a seven-channel microphone. This allowed for eye rotation on a two-dimensional flat screen and body movement towards the person who is speaking. This framework provides a representation of the robot’s attentional gaze that leverages both visual and auditory cues, resulting in more natural and responsive coordinated eye- head gaze movements of the social robot. The potential benefits include improved communication, increased engagement, and a stronger sense of connection with the robot.

It is doubtful that animals have perfect inverse models of their limbs (e.g., what muscle contraction must be applied to every joint to reach a particular location in space). However, in robot control, moving an arm’s end-effector to a target position or along a target trajectory requires accurate forward and inverse models. Here we show that by learning the transition (forward) model from interaction, we can use it to drive the learning of an amortized policy. Hence, we revisit policy optimization in relation to the deep active inference framework and describe a modular neural network architecture that simultaneously learns the system dynamics from prediction errors and the stochastic policy that generates suitable continuous control commands to reach a desired reference position. We evaluated the model by comparing it against the baseline of a linear quadratic regulator, and conclude with additional steps to take toward human-like motor control.KeywordsContinuous neural controlPolicy optimizationActive inference