Anticipatory Human-Machine Interaction

The development of anticipatory user interfaces is a key issue in human-centred computing. A major challenge of any anticipatory interface is to identify relevant information about the users including their environment that may be used by the system to prepare or adjust for future interactions.

In our very early work, we followed the paradigm of the so-called anticipation feedback loop [1] to improve the comprehensibility of multimodal presentation. The overall idea was to simulate how a user would perceive the output generated by a system and to adjust if necessary. For example, by parsing an utterance planned by the system, the generator made sure that the utterance did not contain unintended structural ambiguities. The implicit assumption was that we can predict the user’s understanding processes from the system’s understanding processes.

Our more recent work focuses on the paradigm of Socially-Aware Interfaces that adjust their behaviours based on an analysis of implicitly provided user feedback, such as social and affective signals, as an integral component of an anticipatory mechanism. For example, we investigated to what extent user engagement may be automatically detected from the user’s gestures, postures, facial expressions and speech to adapt a robot’s communicative style if necessary [2]. This approach is based on a direct observation of the user as opposed to a simulation. It bears the challenge that human expressions do not follow the precise mechanism of a machine, but are tainted with a high amount of variability, uncertainty and ambiguity.

The vision for the next generation of intelligent user interfaces (UIs) is to seamlessly assist users during everyday tasks, similar to how humans interact with one another. This is in stark contrast to current UIs, which are primarily reactive and lack the ability to understand or anticipate user actions, intentions, or goals. One fundamental requirement for developing systems that can anticipate human behaviour is the ability to sense and model human attention. While human attention is key to many tasks in human-computer interaction (HCI), such as predicting interruptibility, boredom, or confusion, prior works have mostly modelled attention through proxies. Proxies such as user interactions or self-reports cannot fully capture the temporal, fine-grained dynamics of human attention.

One goal in my research is to leverage either sensed [1] or computational models of human attention for tasks such as predicting user interruptibility, modelling task interleaving, or in the context of work and productivity. Prior research has shown that, e.g. users can be in mental states that make them more likely to be interrupted [2]. My goal is to anticipate or predict when such mental states are likely to occur either using models of attention only, or in a multimodal setting (e.g. leveraging user interactions as well).

Effective anticipatory human-machine interaction (aHMI) requires an accurate model of the human agent. To develop such models, cognitive architectures provide a promising starting point, as such architectures have long been used to develop detailed models of the human cognitive system at a level of abstraction that matches well to aHMI.

A direct example of such an approach is the Slimstampen system for fact learning ([3]; http://www.slimstampen.nl). The core of this system is a model of human memory taken from the cognitive architecture ACT-R (Anderson, 2007), which estimates the memory activation level of each fact that needs to be learned. Based on this information, the system anticipates the next fact that should be presented, thereby optimizing the scheduling of facts (e.g., present a fact again just before it will be forgotten). As a result, users can learn 10% more facts in the same amount of time as compared to standard methods such as flash-card learning.

As a second example, we used ACT-R to develop a model that explains under what circumstances interruptions are more or less disruptive [1]. Based on the detailed account of this model, we then built an interruption management system that measures pupil size to anticipate the best moments for interruptions [2]. By scheduling interruptions at opportune moments in the task, users finished the main task in 10% less time – while being interrupted just as often.

Together these examples show how the use of detailed cognitive models of the human mind can lead to very effective aHMI.

Theory of Mind refers to humans’ ability to infer and represent mental states of others. Such mental states include, for example, beliefs, desires and intentions. Computational Theory of Mind tries to build a model capable of inferring such states [1, 2]. As we want to infer mental states of other agents, Theory of Mind naturally deals with multi-agents systems, in which we have more than one agent observing or acting in the same environment. Multi-agents systems are quite difficult to deal with because one agent’s action can influence other agents’ actions. In this sense, anticipation plays a central role. By inferring beliefs, intentions and desires we can have cooperative agents capable of learning to assist humans or other agents [3]. Agents capable of anticipating others’ behaviour can plan and act accordingly, avoiding mistakes or coordination errors. In principle, this strategy can be applied in both cooperative and competitive tasks or games. Multi-agent AI research has seen most success in competitive two-player zero-sum settings, like chess or Go. However, most real world interactions are cooperative, therefore applying Theory of Mind to improve human-AI cooperation, AI-AI cooperation or even human-human cooperation certainly represents one of the next big challenges in the field.

In human-computer-interaction the action of the technical system can be viewed as feedback to the prior action of the user. Such feedback can fulfil a number of different functions, e.g. signalling that a user’s action has been registered by the technical system, execution of a predefined function, actively provide task-related information, providing motivational feedback, or evaluate a user’s decision. In the neurosciences feedback is inextricably linked with reward rooted in traditional reinforcement-based theories that were developed based on animal studies that applied reward and/or punishment for the animal to learn. This research has even coined the term reward system of the brain consisting of midbrain dopaminergic areas (substantia nigra/VTA) and main projection areas, e.g. the striatum and prefrontal cortex. Following this tradition, the vast majority of studies on the neural correlates of feedback processing in humans have been designed as reinforcement experiments using (monetary) rewards. Consequently, any findings of activation of the brain’s “reward” system has almost exclusively been attributed to mechanisms of reward processing (see Wolff et al., 2020 for references). When we entered the human-computer-interaction domain, we started with the most basic function of feedback, i.e. to signal that a users decision (button press) has been registered by the technical system to study effects of delayed and omitted system response and adaptive processes in the user’s brain towards delayed system responses (Kohrs et al. 2012, 2016). To our surprise, we showed the reward system of the brain to be specifically involved in the processing of such registering feedback. A consistent interpretation of these findings relates to the user making predictions about the feedback of the technical system independent of any reward. For me, this opened up a novel neuroscience perspective on feedback processing in human-computer-interaction because humans anticipate reactions of interactions partners and probably assume such skills to be present in intelligent technical systems, too. We are currently interested in understanding how a technical system can support a user in the course of learning. As use case, we designed a multidimensional category learning paradigm where humans learn the conjunction of two rules by trial-and-error within a session of 180 trials. We observed that many participants persistently stick to a strategy that only considers one to the two rules. In order to prevent the consolidation of such an incorrect strategy, we implemented a heuristic that anticipates such behaviour based on the decision history of the learners. This heuristic then triggers a feedback intervention providing helpful information towards the solution of the learning problem. The results of this intervention in terms of subsequent performance and subjective ratings by the learner how helpful they perceived the intervention was a key motivation to pursue the topic of anticipatory human-computer-interaction: Even though the intervention was objectively helpful, many of the subjects did not rate the support very highly even if they improved in performance. We are now working towards understanding the reasons for this adverse effects of feedback intervention. A simple explanation could be that learners did not like to be interrupted at all or at a later moment or that the mode of neural processing was not compatible with the need to reflect on the previous performance. Whatsoever, my future goal is to include the patterns of brain activity during the course of learning into the process of anticipating the need and the best moment for supportive feedback intervention and to learn from the impact of an intervention on behavioural, psycho- and neuro-physiological data for anticipating the consequences of a proactive intervention of a technical system.

Even after three decades of research on human-computer interaction (HCI), current general-purpose user interfaces (UI) still lack the ability to attribute mental states to their users, i.e. they fail to understand users’ intentions and needs and to anticipate their actions. This drastically restricts their interactive capabilities. We aim to establish the scientific foundations for a new generation of user interfaces that pro-actively adapt to users’ future input actions by monitoring their attention and predicting their interaction intentions – thereby significantly improving the naturalness, efficiency, and user experience of the interactions. Realising this vision of anticipatory human-computer interaction requires groundbreaking advances in everyday sensing of user attention from eye and brain activity. We further require methods to predict entangled user intentions and forecast interactive behaviour with fine temporal granularity during interactions in everyday stationary and mobile settings. Finally, we require fundamental interaction paradigms that enable anticipatory UIs to pro-actively adapt to users’ attention and intentions in a mindful way. We have identified four challenging application areas where these new capabilities have particular relevance: 1) mobile information retrieval, 2) intelligent notification management, 3) Autism diagnosis and monitoring, and 4) computer-based training. Anticipatory human-computer interaction offers a strong complement to existing UI paradigms that only react to user input post-hoc. If successful, work on this topic will deliver the first important building blocks for implementing Theory of Mind in general-purpose UIs. As such, anticipatory HCI has the potential to drastically improve the billions of interactions we perform with computers every day, to trigger a wide range of follow-up research in HCI as well as adjacent areas within and outside computer science, and to act as a key technical enabler for new applications, e.g. in healthcare and education.

My research focuses on AI-Driven personalization, an interdisciplinary field at the intersection of Artificial Intelligence (AI), Human Computer Interaction (HCI) and Cognitive Science. The general objective of this field is to support effective AI-human collaboration by enabling AI systems to personalise their actions and behaviours to the specific needs, states, preferences and abilities of their individual user. Examples of active research in this field include recommender systems, intelligent-tutoring systems, conversational agents, and affect-aware systems.

To provide personalised interaction, an AI system needs to have a personalization loop in which it acquires a model of its user by inferring relevant user properties (see User Model in Figure 1, right) from available interaction behaviors (see Input Sources in Figure 1, bottom) and decides how to personalise the interaction accordingly (see Forms of Personalization in Figure 1, top). In my research I have been investigating AI-driven personalization in the context of Intelligent Tutoring Systems (ITS), intelligent help agents for complex interfaces, and personalized support to visualization processing.

My work in ITS revolves around the objective of moving beyond systems designed to provide support for problem solving, by modeling and adapting to students meta-cognitive abilities and by targeting educational activities that are more open-ended in nature, such as learning from interactive simulations or playing educational games, e.g., [2, 3]. I have done extensive work on using eye-tracking data to inform user modelling, showing that information from eye-tracking can help predict both user short-term states (e.g. confusion, affective valence, learning) as well as long-term cognitive abilities (e.g. perceptual skills, reading proficiency) and traits (e.g. personality). A key application of eye-tracking for user modelling is in driving personalization for tasks that are mainly perceptual in nature, such as processing data visualisations, which has been the focus of my research on user-adaptive visualisations [1]. The key idea underlying this research is to complement traditional approaches to visualisation design based on properties of the target data and tasks by providing real-time support to visualisation processing driven by relevant users states (e.g. attention, confusion) and abilities (e.g. levels of visual literacy, perceptual speed, visual memory etc.).

My most recent research proposes the concept of personalised Explainable AI. This work extends the recent interest in AI systems that can explain their inner workings to their end-users with the idea that these explanations should be part of the personalisation loop (see Figure 1, top), namely they should be dynamically personalised to factors including context, task criticality and user differences (e.g., cognitive abilities and transient states like confusion or cognitive load).

A distinction that needs to be made to understand how AI-driven personalization relates to Anticipatory Human-Machine Interaction (AHMI) is what triggers personalization (often referred to as adaptation in the literature), namely whether the AI agent is personalising to a current user state that it has detected/assessed/perceived or to a user state that is predicted to be happening at some point in the future. When the personalization is driven by the current user state, it is reactive in nature⁵⁵5Note that this reactive personalization is different, and more sophisticated, that the reactivity embedded in traditional interfaces, where each interface action (e.g. a button press) is associate by design to a specific system’s response, which does not depend on the user who performed the action.. Personalization driven by a predicted future state of the user can be viewed as preventive/proactive/anticipatory. I see the latter as being the key aspect of AHMI. I will illustrate this distinction with examples based on some of my work.

An example of reactive personalization in my ITS research comes from my work on ITS for supporting students in acquiring meta-cognitive skills relevant for learning. Specifically, in [2] we focused on the meta-cognitive skill known as self-explanation, i.e. one’s tendency to explain and elaborate instructional material to themselves when studying. There is ample evidence in Cognitive Psychology that this skill is conducive to better learning, and that students who lack the skill can acquire it if they undergo instructional sessions during which a tutor prompts them to self-explain if they do not do it. During these sessions, students are asked to self-explain aloud, just so that the tutor can detect if they are self-explaining or not. However, self-explanation usually happens in one’s head. In [2], we worked on an ITS that assesses if a student self-explains in their head while studying given material, based on their gaze patterns and interface actions. The ITS then reacts if it detects that self-explanation did not happen for a specific aspect of the instructional material, by generating prompts to encourage and guide the student to self-explain that material.

An example of anticipatory personalization in my ITS research comes from my work on ITS for supporting students learning through explorations in open-ended-learning environments. The capability to explore effectively is relevant to many tasks involving interactive systems, but not all users possess this ability. I have been investigating ITS that can provide personalized real-time support to users who are having difficulties with an exploratory task, while interfering as little as possible with the unconstrained nature of the interaction. Building such ITS is challenging because it requires having a user model that can predict the effectiveness of open-ended behaviours for which there is often no formal definition of correctness. Our approach relies on clustering and association rule mining to discover from data the user exploratory behaviours conducive or detrimental to learning. Supervised machine learning is applied to the resulting clusters to derive classifiers that predict in real-time whether a user will learn from the interaction and, if not, it anticipates how the user behavior should change to increase their learning. Based on this aanticipation, the ITS then generates hints designed to foster the desired change in behavior [3].

The distinction between reactive and anticipatory personalization is one that has not yet been made in the literature, but it is important to think about the unique advantages and challenges that anticipatory personalization brings to AI-Human collaboration.

Systems built with the purpose of human-machine interaction do not necessarily need to behave optimally in each circumstance when interacting with humans. However, they should have the ability to provide socially acceptable explanations for their decisions. They might generate such explanations by having an internal representation of the current state of the human mind (theory of mind). This representation might allow the system to predict future events and possibly prepare by intervening with some action (e.g. to avoid catastrophic outcomes or to improve the human’s situation).

The internal representation of predicting and deciding on which action to choose can happen on various levels. An important requirement however, is that the system should have the ability to provide socially acceptable explanations for their decisions. More specifically, their explanations should have some basic reasons of support (attributive), explain why a conclusion is supported in contrast to opposing conclusions (contrastive) and provide information that guides on how to act following the conclusion (actionable). Cognitive argumentation fulfills these requirements. In cognitive argumentation, having its roots in computational argumentation, the dialectic argumentation process is assumed to be heavily guided by biases or heuristics [1, 2]. Initially, a root argument for a certain position is built (e.g. based on the awareness in the current situation), (strong) counterarguments are searched (e.g. factual information contracts the argument), which the root argument defends, and is either replaced by stronger arguments or extended with the defense. This process seems similar to sensemaking, in which humans generate mental models to explain events, question these models, and reframe or replace these models by more adequate ones.

In the context of human-machine interaction anticipatory thinking can be implemented through future oriented cognitive argumentation by generating future arguments for or against carrying out a possible future action (of the human or machine) that leads to the most desired state.

My research concerns the development of visualization methods to gain insights from large quantities of scientific data (typically acquired from extensive experiments and simulations). Interactive approaches enable explorative user analysis of rich and complex data, but respective visualization systems generally exhibit numerous and degrees of freedom regarding how information is presented and explored.

In many of my previous works, techniques and systems are configured in a data-driven way, e.g., which time steps [1] or spatio-temporal subsets [2] to convey and how to layout large numbers of graphical tiles depicting members of an ensemble [3] (Figure 2). For this, these approaches generally optimize application-specific metrics or employ self-supervised learning. For example, metrics quantify how well dynamics in a full time series are captured by a subset [1] or the similarity of ensemble members positioned nearby in a grid layout [3]), and learning can be used identify unusual behavior [2].

Conceptually, these approaches optimize the transformation of the data into an (interactive) visualization, but largely neglect – or only indirectly consider – the human decision-making process operating on this basis. Explicitly incorporating this aspect, the anticipation of user performance for a given (configuration of) visualization system allows to explicitly optimize the visual representation and interface for a user to gain insights more quickly and/or more accurately.

A major challenge for research in artificial intelligence is to develop systems that can infer the goals, beliefs, and intentions of others (i.e., systems that have theory of mind, ToM) (Nguyen & Gonzalez, 2021). To generate relevant Human-Machine collaborations, one needs to create computational representations of human behavior. In the past decades cognitive models have advanced towards more accurately representing and replicating human cognitive processes. In particular, we have been involved in generating models that can explain how humans make decisions, learn, and adapt to improve their choices in dynamic tasks. We rely on Instance-Based Learning Theory (Gonzalez, Lerch & Lebiere, 2003; Nguyen, Phan, & Gonzalez, 2022) to construct models of human dynamic decisions in a large number of situations. Our applications for anticipatory Human-Machine Collaborations include but are not limited to: Applications in cybersecurity, replicating attacker’s behaviors that can be used to inform defenders; and replicating end-users’ phishing behaviors that can help prevent successful attacks. We have also used these models in dynamic control tasks and scenarios for search and rescue tasks, where predicting the state of the mind of teammates’ decisions is essential.

Research in human-robot interactions (HRI) usually focuses on either cognitive interactions or physical interactions. However, characteristics of HRI in real-life scenarios often involve both information exchanges on the cognitive level and force exchanges on the physical level. Specifically, the human agent and the robotic agent need to infer each other’s intentions, predict each other’s actions, and act adaptively in the given environment [1]. In such dynamic interactions, the robotic agent also needs to anticipate in order to support the human agent with flexibility. Therefore, it is crucial to create unified models of HRI that can capture the dynamic exchanges on both cognitive and physical levels while taking the environment into account (see [2], for an example of such a unified model of a single agent). Our work aims to provide a conceptual framework with possibilities to connect models of physical and cognitive HRI. We also hope to potentially apply our framework to an anticipatory robotic agent that provides the human agent with different levels of assistance.

Humans have the astounding ability to attribute mental states to their interaction partners and reason about their behaviour – so-called Theory of Mind (ToM). Similarly, humans also attribute a mind and human abilities to AI systems (like intelligence and learning). The attribution of a human-like mind to the AI system puts the focus on the AI systems, thereby loosing sight of the human decisions about who and what gets counted when building them. The thought processes, intents and decisions can be surfaced by redirecting Theory of Mind away from the AI system and towards the AI experts who built the system. The redirection necessitate anticipatory human-machine interaction in order to (1) infer the mental states of the AI experts and (2) predict the explanations needed by the user of the AI system. In previous work, we have shown that intents – a mental state with explanatory power – can be predicted from mouse and keyboard action in a text formatting task[1].

At this Dagstuhl seminar, I have taken the position that the transportation domain can benefit from anticipatory models for human-machine interaction. One specific domain that I have investigated over the last few years is semi-automated driving. Here, the vehicle takes over some components of the driving task (such as lateral or longitudinal control) under specific conditions (or operational design domains; such as highway roads). However, typically, the vehicle cannot drive under all circumstances, and some human intervention is occasionally needed. The default model in the literature is to look at what the minimum time is that humans need between a warning and the time where they need to act. There seems to be consensus that somewhere between 5 to 8 seconds is a minimum time, and indeed the large majority of the literature has looked at response time situations of around 2 seconds. However, as I have argued elsewhere in more detail [2], although it is beneficial for research to understand what the minimum response interval needs to be, there can be many situations where it is useful to think more ahead and maybe have “pre-alerts” or forewarnings. One reason might be that as automated vehicles gain more capacity, human drivers might not always pay as much attention to the as they would under non-automated conditions, and start performing other tasks. Working on non-driving related activity is something that some drivers already do under regular (non-automated) driving, but that they want to do under higher levels of automation. In these situations, where attention has been to another task or activity than the driving task, one cannot assume that the attention of the driver has been on the road and that they had sufficient situation awareness to act in future (emergency) situations. Moreover, they might not even process the alert itself instantly. So, instead of thinking about fast task switches within, say, 5-8 seconds, one can instead think of how the process of returning attention from a non-driving activity to driving unfolds. I have proposed a conceptual model to capture this [2], of which the first empirical tests have taken place [2]. Now, the next step is to develop models that try to analyze human state and act accordingly, to interrupt at opportune moments. Such models need to take into account the user state, but also the environment and the driving scenario (e.g., is it a situation where there are many alternative actions, or is the intended action quickly clear from the context?). Personally, I would hope that techniques such as Hidden Markov Models can be effective in capturing human states and its dynamics [3]. Without anticipatory models, interruptions and alerts might come at less opportune moments, which can lead to increased stress or workload, and leave the driver in no capacity to act quickly and safely. Although I have specified these principles for driving, I anticipate that they also apply to other traffic domain such as railroads, ships, and airplanes.

Current AI-assistants are only able to help us when we can precisely specify what we want – that is, the goal and the problem setting. However, that arguably is not when we most need assistance, and effectively the current assistants require us to program them, just with a different kind of an interface such as giving precise spoken commands in restricted language. Instead, we need assistance the most we do not yet know precisely what we want, because the goal may be tacit, unclear or evolving. I am developing AI-assistance principles for situations such as drug design or scientific research, where the goal initially is unclear and evolves during the process.

Such assistants need to anticipate their users to be helpful to them: both their actions, some of which they can assume to arise from the user’s planning towards their goals, but also their goals and the development of the goals, and furthermore how the users would understand the assistant’s actions. This is particularly difficult because of the nature of the design-type tasks, where each task is different and laborious, resulting in scarcity of data for learning the goals. In particular, this culminates in “zero-shot assistance” for design and decision making, which I believe requires (i) probabilistic modelling to handle the uncertainty arising from scarcity of data and a large number of potential explanations of behaviour, (ii) nested POMDP-type modelling to capture a “theory of mind” of the user, to both infer the goal by reverse-engineering the behaviour and to anticipate how the user would understand the assistant’s action, and (iii) novel computational approximations to actually do the necessary inferences.

As a result, we hope to have better tools for solving the grand challenge problems humanity currently is facing.

In psycholinguistics, surprise is a key concept. It measures how predictable the next word in a sentence is, given the prior context of the word. Surprise has been shown to correlate with behavioral measures such as cloze probabilities, reading times [1], or EEG signatures such as the N400. Mathematically, surprise is the negative log of the conditions probability of word $n+$ given words $1\mathrm{\dots }n$, where a language model is typically used to estimate this probability. Recent work in computational cognitive modeling has used surprise as a component in a more comprehensive cognitive model. For example, the Neural Attention Tradeoff model of [2] models the allocation of human attending during reading: it predicts which words are skipped by the reader, and which ones are fixated. Skipping in the model depends on surprise, but also on aspects of the reading task (e.g., reading for pleasure, proofreading, reading to answer a question).

There is an intuitive connection between anticipation and surprise. Assume an agent predicts the next word, and based on that decides whether to skip it or not. If that decision turns out to be incorrect, then the agent will be surprised, which is something we can measure behaviorally. Surprise can therefore be conceptualized as unsuccessful anticipation. It is worth exploring this idea in the present seminar, and in particular discuss to what extend surprise generalizes to domains other than language processing.

Cognitive scientists have not only studied language processing in single readers and speakers, but have also investigated pairs of speakers (dyads) that communicate and interact to solve a shared task. In such a situation, dialogue participants are known to align with each other. This means that their linguistic productions become more similar across a whole range of dimensions: they use similar words, similar syntactic structures, adapt their speech rate and even their pronunciation. At the same time, they also align non-verbal behavior such as gaze, posture, and gesture. The degree of alignment typically increases over the course of the dialogue, and alignment is correlated with task success. A classical conceptualization of this process in the Interactive Alignment Model [3]. More recent work has refined our understanding of alignment: for instance, using a collaborate search task, Coco et al. (2018) show that the relationship between alignment and task success is not absolute, but depends on how interactive the task is.

Alignment is relevant in the context of anticipation. A participant that aligns with their collaborator predicts the future behavior of that collaborator and decides to make their own behavior more similar. Alignment is therefore a form of long-term anticipation and, like anticipation more generally, it makes collaborating on a task more efficient. This is another observation that could feed into the discussion at this seminar, where questions to be explored include: Does the relationship between alignment and anticipation generalize to non-verbal tasks? Does alignment also make human-machine interaction more efficient? Should machines align with humans? Are humans able to align with machine?

The interaction of two pilots in the cockpit benefits from their ability to anticipate each other’s behaviour. Single pilot operations will be characterised by increased interaction with higher levels of automation instead of with another human pilot. By modelling the pilot and their mental representation of the automation, the automation can anticipate the pilot’s behaviour and adapt its own behaviour accordingly to stay in line with the pilot’s anticipations [1]. Cognitive architectures like ACT-R have been used to model pilots [2, 3]. In particular, instance-based learning in ACT-R has been shown to be effective in learning mental representations through observations of behaviour sequences and in modelling intuitive associative processes underlying anticipation. We hypothesise that instance-based ACT-R models of pilots, for example based on training on flight recorder data, can enable anticipatory human machine interaction in the cockpit during single pilot operations.

The visual analytics community has made significant strides in developing systems that facilitate the interplay between humans and machines in exploratory data analysis and sensemaking [1]. These advances have enabled us to learn from user interactions and uncover their analytic goals (e.g., [2, 3]). Moreover, they have set up the foundation for creating visual analytic systems that anticipate and assist or guide users during data exploration. Providing such guidance will likely become more critical as datasets grow in size and complexity, overwhelming the limited screen real estate and human cognitive power. My team and I use interdisciplinary approaches to identify optimal data representations for effective decision-making and design human-in-the-loop visual analytics interfaces more attuned to domain scientists’ exploration goals. Still, some open anticipatory challenges and considerations may be unique to the visual analytics setting. In particular, the role of the AI may depend on the task and include but is not limited to: (1) the AI can be a teammate, working with the human to collaboratively solve a problem by either performing the same task concurrently or separate complementary sub-tasks. (2) the AI may be an assistant, observing actions and taking preparatory steps such as prefetching and precomputation. (3) The AI may serve as a learner, proactively observing actions to create a knowledge graph or process script. Moreover, the user’s tasks will likely evolve. Thus, anticipation should happen at a global level (e.g., deciding the role the AI needs to assume based on the observed state) and at a local level (e.g., deciding which data points to prefetch). Finally, the community should address ethical considerations for how AI interventions might influence the user’s analysis. For example, we need to preemptively consider how to track biases in the data shown to the user.

For gaining mutual understanding between a human and an intelligent agent the intelligent agent needs to develop a model of the human partner and the environment as well as of the task. Depending on the type of interaction and purpose of the interaction different requirements arise for a fluent and efficient collaboration. For more complex interactions the assumption of a shared representation about the individual goals, the task and situation understanding is essential. The assumed shared representation of technical system and human might differ. Direct and indirect communication is necessary to provide information about the individual shared representations. Especially for physical Interactions with technical Interactions a lot of different cues make sure that e.g. a handing over procedure of an object will be well coordinated in time, location and type. What types of cognitive models are needed to address these issues? Since anticipation is the ability to prepare in time for problems and opportunities –it is functional in form of preparing for future even unlikely but critical events and not simply predicting what might happen. So context and dynamical representations would be relevant. It could be useful to focus the (cognitive) model rather on the relevant components of e.g. (dynamic) situation understanding and decision making for being able to trace behavior and cognitive state updating rather than capturing all aspects of cognition (Klaproth et al. 2021, Scharfe et al 2019). For modelling anticipatory ability in a physical HRI further fine-grained social cues must be detected and also produced for realizing joint action. One of the big challenges I see here is to find good ways for information integration and hybrid modelling approaches that are able to meet all the different requirements the individual tasks pose.

Details are given in [1] and [2].

Intention-related neural signals may be exploited by different human-machine interaction systems. In particular, brain-computer interfaces (BCIs) could be used to restore communication functions and facilitate motor control in paralysed users. BCIs record the brain activity of the user, detect the user’s intended action and send the corresponding execution command to an artificial effector system, e.g. computer cursor, prosthetic device, etc. Non-invasive brain recording modalities such as electroencephalography benefit from a fine temporal resolution for the signal acquisition which facilities the anticipation of different cognitive or motor states in a timely manner.

In my talk, I presented an approach that involves the simultaneous acquisition of multimodal data coming from both neural and behavioral information [1], with the purpose of understanding the underlying mechanisms that link perception to action within the framework of brain-computer interfaces. Specifically, in the context of grasping movements, I showed the importance of disentangling at a neural and behavioral level between the properties of the objects (such as shape and size) and the properties of the movement (such as type of grasp, number of fingers involved) [2]. Moreover, I showed that this disentanglement is similarly observed across modalities in separate stages going from perception, observation to action. Furthermore, I opened the conversation about the influence of the goal and feedback on the perception of action in anticipatory human-machine interaction cases involving different levels of attention to the action.

In many machine learning areas, the goal for a model is to solve a specific task faster or better than humans can. However, there are many areas where humans cannot be replaced by an artificial intelligence (AI) system for different reasons, e.g., because humans might not want to be replaced but only seek assistance from an AI, or if the task to solve necessarily includes a human or at least human information. As a consequence, there is an increasing need for AI systems that collaborate with humans, thereby enable them to solve a certain task more efficiently or to achieve a certain goal faster or more easily. For an AI system to collaborate with humans, it is important that the AI understands which information the user is perceiving and processing, as it then can use this information to anticipate user intents. Once the AI system identified user intents it can, depending on the environment and task, anticipate the user’s intent, generate results, or perform actions to assist the user. In order to understand which information a human is perceiving and processing it needs to receive information about the user. We specifically focus on using behavioural data as a source of information about the user. The advantage of behavioural data is that the user does not have to explicitly provide it but the AI can automatically obtain user information through their interaction with an application or through external devices like an eye-tracker or using electroencephalography (EEG). [1] One of the main challenges of using implicit behavioural data is the understanding of such data and the subsequent extraction of task relevant information. We plan to develop methods to extract the most relevant information, especially concerning gaze data, e.g., how task relevant a specific fixation is. Since collecting large datasets of such attentive behavioural data is time-consuming and expensive, we will propose methods that allow to generate synthetic datasets and model human behaviour to train intelligent systems.

My specialization is Cognitive Modeling using Cognitive Architectures. I have been involved in ACT-R for a long time, and have contributed mechanisms for production compilation, multitasking and time perception. More recently, I have developed my own variant on ACT-R called PRIMs [1]. The motivation for this was the fact that ACT-R and many cognitive architectures are focused on individual tasks, whereas humans have a history of prior experience that they can use to perform new tasks.

To support models that can perform multiple tasks, I defined an additional level of abstraction between production rules and (task) goals: the level of skills. The idea is that if people have to perform a novel task, they try to mobilize the skills necessary for that task from their available repertoire [2]. Only when they miss particular skills they need to do more extensive learning in order to acquire the missing pieces.

An application area of the skill idea is education, in particular cognitive tutors. Using unsupervised learning methods, we can extract the skill set for a particular domain. Then, give a student’s performance on assignments, we can determine which skills the student has mastered, and which not. Based on that, the optimal material or assignment can be selected for the student. In addition, it can be used to provide feedback to both the learner and the teacher.

The purpose of this group was to discuss the suitability of rule-based models of anticipation compared to data-driven models. We start with a definition of these terms. Traditional examples of rule-based models include the ACT-R cognitive architecture. However, even ACT-R contains a learning component, e.g., the activation levels are trainable. Other rule-based models includes most models used for planning, and more generally logic-based models. The group felt that some tasks are specified symbolically (e.g., in robotics, where the starting state and the end state of an action is given), and this might bias any modeling effort towards rule-based models. Also, it is important to bear in mind that even rule-driven models are built based on data, except that the model designer comes up with the abstractions (e.g., based on experimental results), and formulates these abstractions in the form of rules.

The prototypical data-driven models are deep-learning models, which are black-box and contain very little internal structure, and are trained on massive datasets. Models with a simple structure are also included, e.g., models that learn state transitions (HMMs, POMDPs). Instance-based learning is another example. It needs to be taken into account, though, that some data-driven models have internal structure (trees, graphs), for example graph convolutional neural networks.

The group also considered hybrid models and a third model type that combines both data-drive and rule-based models. Hybrid approaches are prima facie plausibility, for example because humans probably learn in a hybrid way, combining pre-specified rules (instincts) with data obtained from the environment. Language acquisition is a potential example: the learning mechanism seems to be innate, but external data is required to trigger the acquisition of an individual language. So it seems clear that you can’t learn without data, but also not without (possibly pre-defined) structure. A classic example of a hybrid model in machine learning are Bayes Nets, which combine pre-specified (graph) structure with probabilities learned from data.

The group then switched to the discussion of Theory of Mind (TOM) and its relevance to modeling anticipation. As to the definition of TOM, we can think of it as a nested model, i.e,, a model within a model. More specifically, having a TOM means having a model of another agent that accounts for their knowledge, beliefs, and intentions. This raises the question of what TOM is useful for and which behaviors we can’t do without TOM. Traditionally, non-literal behavior is thought to require TOM: lying, irony, sarcasm, empathy. The group then examined the question whether TOM can be learned in a data-driven way. This is probably possible in specific instances, but generalization would require a rule-based approach. The group also believed that TOM would not be needed for anticipation in the general case; limited, task-specific anticipation behavior could be learned from data. The group also noted that a higher level of abstraction is required to generalize anticipation across tasks, and TOM may be required for that. The link between (impaired) TOM and autism was noted.

There were other issues that seemed important to the group, but were not discussed in detail because of lack of time. This includes the question of how a simulation is different from a model. The group felt that a simulation required models of both the user and their environment. A simulation could also be used for reasoning, e.g., to simulate what a user does in a given environment and therefore reason about what the user would end up doing. Another important topic that the group touched upon was explainability. Humans constantly generate explanations for their behavior and their actions, they tell stories that help them to make sense of the world and to interpret other people and their actions. Another issue the group touched upon was how TOM relates to sense making. Sense making is a way of model criticism, of challenging a model, adapting and modifying it. It was also pointed out that Inference is possible without probabilities, e.g., in Approximate Bayesian Computation (ABC). Another important issue seems to be causality. Can data-driven models be causal? The group felt that this was possible in certain cases, a good example are Bayes Nets.

Finally, the group discussed evaluation. This is potentially an issue with respect to theory of mind: how can you tell that a system has TOM? It may be possible to use tests designed in developmental psychology to test TOM in human participants, e.g., the false belief task (Sally-Anne task). These tests would have to be adapted for a human-machine interaction setting. This raises the wider question of how to evaluate anticipation. How do we tell if a system anticipates successfully? One approach would be to rely on task-based evaluation, i.e., just measure task success. This gives us an indirect way of evaluating anticipation. However, a direct evaluation measure (“anticipation accuracy”) would be desirable, but this remains an issue for future work. Furthermore, the group noted that data-driven models cannot be trained without an evaluation measure that informs the design of the loss function.