Direction des Relations Internationales

CuriousTECH Inria Associate Team

Curiosity-Driven Learning Across the Lifespan 

  1. Research themes on digital science:  AI ; Human-Computer Interaction ; Virtual Reality ; Cognition
  2. Other research themes and application areas:        Cognitive Science ; Education ; Health ; Aging

Participants

Flowers Research Group (centre Inria de l’université de Bordeaux)Rania ABDELGHANI | PhD candidate | Master of Science | National Institute for Research in Computer Science and Control, Le Chesnay | INRIA | FLOWERS - Flowing Epigenetic Robots and Systems Research TeamMaxime ADOLPHE - Étudiant doctorant - Inria x onepoint | LinkedInRania ABDELGHANI | PhD candidate | Master of Science | National Institute for Research in Computer Science and Control, Le Chesnay | INRIA | FLOWERS - Flowing Epigenetic Robots and Systems Research TeamMarion Pech - Chercheur postdoctoral - Inria | LinkedIn9 profils pour “Desvaux” | LinkedIn

University of Waterloo (Canada)

*PI of CURIOUS Tech team

Other Participants 

Other Collaborators

Objectives  & Purposes (2023-2025)

To update, curiosity research in humans still is in its infancy. Only a few curiosity studies exist that address the fundamental questions about curiosity, such as its neurophysiological bases or its evidence-based impact in learning activities with respect to various individual's characteristics or real-life settings. The science of curiosity deserves more investigations and research efforts that explore the strengths and weaknesses of curiosity-based interventions across individuals, tasks and situations, especially in learning fields.

More importantly, current curiosity studies are largely done in disciplinary silos, with few if at all, cross-disciplinary collaborations. Yet, there are many examples that show how multidisciplinary approaches can produce qualitative leaps in scientific progress, especially for understanding complex systems or constructs, such as stochastic modelling of human behaviors in high-risk human-machine systems (avionics and nuclear industries).

Leveraging our recent results on curiosity benefits for learning in both children [3, 4, 5] and older adults [6, 7], our proposed associate team aims to develop an original, cross-disciplinary approach, joining together two perspectives:

1) The fundamental study of curiosity-driven learning across life-span (in children, young adults and older adults) and

2) The study of how new (re)educational technologies, using both curiosity-related models and artificial intelligence techniques [3, 8, 9], can personalize learning sequences for each individual, maximizing curiosity and learning efficiency in real world contexts.

Our proposed research will produce new understanding of the role of curiosity in education and healthy aging, through the design and the field assessment of new interactive educational technologies or health-related technologies. Beyond academic contributions, we expect our findings to inform the broader societal challenges inherent to the School of the 21st Century, ranging from helping children (and their teachers) to develop cross-domain skills for learning such as curiosity and meta-cognition, while improving inclusivity in schools (learners with disabilities, especially cognitive disabilities) as well as promoting lifelong learning in older adults (successful aging), using cognitive-based research findings.

Another outcome of our joint program is to use applied research to accelerate the transfer of results to industries and public institutions related to education and healthy aging in both countries. The mixed method approach used in our proposed project (user-centered methods, digital technologies, artificial intelligence, and field assessment) will help demonstrate the effectiveness of our developed technology, and facilitate adoption by industry partners and market stakeholders from various education and health care organizations.

Work Program  (2023-2025)

Our investigations will be grounded in the design of (re) educational apps based on cognitive science (i.e., using educational and cognitive aging frameworks) as well as adaptive machine learning algorithms, such as ZPDES algorithm [3]. The apps will target the the learning of academic skills as well as skills required for everyday functioning. To evaluate the apps, we will design and conduct two sets of experiments with multiple diverse pools of participant, covering a large age range and diversity in cognitive functioning and socio-economic status. This setup will allow us to examine the positive role of curiosity for learning in all ages and its protective role as a fruitful approach to promote successful cognitive aging.

The first set of experiments will be dedicated to the fundamental exploration of curiosity processes through experiments analyzing the neurophysiological correlates (e.g. EEG-Electroencephalography and other physiological signals analyzed with statistical machine learning tools) of curiosity and the robustness of curiosity benefits in learning performance for all the ages (children, young and older adults). In particular, the goal will be to find the signature of various levels of curiosity in neurophysiological signals such as EEG, GSR (galvanic skin response) or HR (heart rate). In an original and exploratory way, we will use machine learning techniques from the field of brain-computer interaction (BCI) [9] to attempt to recognize in EEG, GSR and HR various states of curiosity when users participate in curiosity-inducing tasks (e.g., Trivia). Recognizing these signatures of curiosity will then enable us to use such signals to monitor curiosity in real-time and possibly adapt applications on the fly to stimulate curiosity. Then, whether and to what degree curiosity benefits learning performance in children, younger and older adults will be compared. To this end, we will use the most recognized methods of promoting curiosity (the trivia paradigm or the manipulation of information gap or uncertainty) to determine their effects on learning. We will do so by varying the contents and (re)educational objectives for each task (e.g., acquiring new knowledge from the reading of expository texts, and free exploration and navigation in large-scale spaces of virtual environments).

A second set of experiments will focus on the design of (re)educational apps, with the ambition to create new curiosity-driven interactive systems using various digital media (social agents/robots, brain-computer interfaces, etc.). One example of such technologies is a curiosity-based training app that uses a social agent to teach “question-asking skills” in order to foster deep reasoning during reading tasks, increase motivation and encourage effort, and accelerate learning.  Socio-emotive selectivity theory posits that different motivational goals guide cognition as these differ across the lifespan [11] ; and so, central to our investigation is also the question of how to personalize the app to each individual, depending on their age and other personal characteristics.  Finally, we will empirically assess the utility of these apps within laboratory environments as well as field studies in real-world contexts (schools, homes).

Scientific Results (preliminary results) and Future Works

Route Memory Benefit from a curiosity-based exploration of new (virtual) environments

Figure A. First-person view and bird’s-eye view of the three styles of virtual environments. Participants only experienced the environments from a first-person perspective.

Panel a. City                         Panel b. Mall                        Panel c. Park 

Figure B.  From left to right: Condition 1 with Low Uncertainty (1 character); Condition 2 with Medium Uncertainty (3 characters); and Condition 3 with High uncertainty (7 characters)

Une image contenant texte, table Description générée automatiquement

As trait curiosity increased (Fig. C), so did memory performance in the high uncertainty condition, suggesting that children with high levels of curiosity can better recruit cognitive resources within such environments. Children with high compared to low curiosity also had higher feelings of presence during the immersive experience. Importantly, in environments with medium uncertainty, children with low trait curiosity were able to perform as well as those with high curiosity. Results suggest that individual differences in trait curiosity influences route memory in environments with varying levels of uncertainty.

Figure C. Route Memory Score (black circles) and Intrinsic Motivation Score (white circles) in Low-and High-curiosity Groups as a Function of the Three Uncertainty Conditions (Low, Medium and High)

In simple words, we observed that the virtual environments eliciting an optimal curiosity state enabled children to enhance their memory performance, and more particularly for children with a lower propensity for curiosity. Thus, stimulating curiosity through the characteristics of the learning environment is an effective way of helping children who do not spontaneously engage in inquisitive exploration of the environment

However, the benefit from a curiosity-based exploration in a subsequent experiment was cancelled when body movements were limited during exploration.

Embodiment seems to play a critical role in the benefit from curiosity-based exploration for spatial learning.

Figure D. Experiment setup and Memory performance according to both the exploration and the age factors.

Results showed that active navigation where exploration is free, compared to passive viewing during encoding, resulted in a higher accuracy in spatial memory, with the magnitude of this memory enhancement being significantly larger in older than in younger adults. This supports that age-related deficits in spatial memory can be reduced by curious active exploration.

Curiosity states seem to play a protective role against age-related decline in spatial learning.

Future works

For future directions (Year 2), we will carry out experiments similar to those previously carried out with children and young people, but this time with healthy older adults. The rationale is to assess the potential protective effect of curiosity states on age-related memory decline, and to evaluate the role of enactment benefits known to be preserved with aging.

Also, in young adults, we will explore the role of agency in curiosity-driven learning. In particular, we have initiated work on training medical students in neuroanatomy and neurosurgery using novel interactive systems (AR and VR) (PhD’s Work program of M. Poupard with Catie and the Epsilon Lab) (Year 2 and 3).

Designing artificial conversational agents to train children's curiosity during learning,

 a proof of concept through the “Kids Ask” project

Our joint “Kids Ask” project leverages new technologies and proposes a curiosity training based on the awareness of “knowledge gaps”; paying particular attention to self-assessment and how to avoid the “knowledge illusion” trap. In particular, we implement conversational agents capable of stimulating questioning and exploration motivated by the desire to compensate for specific missing information. Our work is notably motivated by the positive effects that this type of technologies has shown on children’s learning strategies, their divergent thinking, as well as for the construction of good social relationships with them.

Design

Kids Ask is an web-based educational platform prototype, that involves interaction with a conversational agent. The platform is designed to help children identify their uncertainties, or “knowledge gaps”, generate questions to compensate for these gaps and, finally, use their questioning skills to explore and acquire new knowledge autonomously by following their curiosity.

To meet this objective, “kids Ask” platform offers three work spaces (Fig. 1):

  1. Curiosity stimulation Space: The agents in this space offer quizzes on different topics and try to make students aware of gaps in their knowledge by encouraging them to report confidence levels in the answers they give. We use this strategy to probe students’ meta-cognition and attempt to arouse their curiosity by making them aware of gaps in their knowledge. Indeed, these strategies allow children to think more deeply about what they know and do not know and thus be able to defer more uncertainties [14]-the first step to adopting curious behaviors.
  2. Curiosity elaboration space: In a second step, our agents aim to help children go one step further than the identification of uncertainty: we want to train them to know how to express these uncertainties and to be able to pursue them by asking the appropriate questions.
    To do this, the agents propose reading-comprehension tasks and specific cues for formulating questions that are said to be “divergent” from the texts. These are questions that bring up new information about the text and ask students to make hypotheses, connections between different ideas, etc. The agents help the students in this exercise by proposing precise linguistic and semantic clues that lead to this divergent thinking. The idea here is to propose “knowledge gaps” to the students and to encourage them to pursue them by formulating high-level questions. By doing this exercise, children could thus acquire new information that they would have looked for on their own, and which is important for a better understanding of the text. We choose to focus on training divergent questioning because it is an ability that induces more curious thinking and involves the use of more cognitive processes.
  3. Curiosity maintenance space: In order to help children stay in curious states, our agents encourage them to mobilize their questioning abilities to explore educational resources available in the platform and perform autonomous, organized, and personalized investigations. The purpose of this space is to train children to take time to evaluate the new information they acquire and ask themselves if they are satisfied with it or if it raises new questions in them. Such exercises help avoid the trap of the “illusion of knowledge” and the “premature” interruption of learning cycles.

Technical implementation of “Kids Ask”

The interface was programmed in JavaScript using the REACT library and was connected to a RESTful API to publish and retrieve the interaction data. The behavior of the agent in terms of selection of the adequate cue(s) to offer was predefined and hand-scripted: it was connected to a database containing the different text resources and every text had a sequence of linguistic and semantic cues linked to it. Depending on the child’s condition, the agent’s automaton composes the dialogue utterances in order to include the appropriate support. We changed the utterances between the questions to avoid repetition in the agent’s dialogue: a replica is not executed if it has been delivered during the previous question.

Our current implementation had no natural language processing methods to condition the agent’s behavior: the recommendation system in the exploration space was only based on an automaton that shows the resources related to the topic of choice of the child. Also, the type of question entered by the child was only assessed later on during the data analysis phase.

Results and current works

The “Kids Ask” platform was tested with four classes from two different French elementary schools; participants were aged between 9 and 10. The results (Fig. 3) suggest that curiosity-driven behaviors such as curiosity-driven questioning and independent explorations can be trained on and enhanced with the help of educational technologies in general and conversational agents more particularly. We also show that these behaviors have a direct positive impact on the learning progress children can make.

In simple words, the “Kids Ask” platform using conversational agent providing curiosity-incentives, was successfully tested with participants aged between 9 and 10.

After these positive results, we worked on the automation of our agents’ curiosity-prompting behaviors in order to facilitate their implementation on larger scale and in different school activities. For this, we are studying the possibility to leverage the advances in the natural language processing field, i.e., ChatGPT3. We recruited 75 4th grade students aged between 9 and 10.5 years old and assigned to one of the three experimental conditions of agent’s design: 1) incentive hand-crafted agent ; 2) the incentive GPT-3-driven agent and 3) the open GPT-3 agent (see agent description in Fig. 4) (Abdelghani et al., 2023a).

Figure 4 : Agents’ behaviors and implementations for the three experimental conditions

In evaluating the cues generated, we noted no offensive output from GPT3 and the three conditions were similar in terms of both the divergence level and the semantic relatedness to the text with what we had before by hand (as evaluated by human annotators) (as illustrated in Fig. 5).

Figure 5 : Left : Semantic relatedness of the cues from GPT3 is similar to what was produced by hand.

Right: Divergence level for the incentive semantic cues was similar for the hand-crafted and the GPT3 versions.

Importantly, the results in terms of learner experience (elicited motivation and cognitive load, Fig. 6) as well as of curiosity behaviors (elicited question asking, Figure 7) have been conclusive by revealing equivalent results in children with GPT3-generated incentives, or even a superiority effect of the “open” condition of GPT3.

Figure 6 : Learner experience measures

Figure 7 : Elicited curiosity-based behavior -Question-Asking performances during the training

The overall results suggest the efficiency of using LLMs to support children in generating more curious questions, using a natural language prompting approach that affords usability by teachers and other users not specialists of AI techniques.

Furthermore, open-ended content may be more suitable for training curious question-asking skills.

Identify-Guess-Seek-Assess (IGSA), a framework for modeling and training children’s curiosity-driven learning through metacognition

Leveraging the hypotheses about the roles of metacognition in organizing curiosity-driven learning, we recently introduced a training paradigm aiming to teach curiosity-driven learning via training of metacognitive skills (Abdelghani et al., 2023b). This training paradigm is based on considering the four basic metacognitive skills of Murayama’s framework [12]: identify (I) a knowledge gap, guess (G) possible answers, seek (S) information to fill the gap, and assess (A) the quality of the information (Fig. 8).

Figure 8 : Curiosity-driven learning framework and link with the metacognitive skills we propose to train as facilitators during our IGSA-based  intervention

Design

In the training paradigm, the relevant metacognitive skills are personified as animated metacognitive characters as follows:

  1. The first character is the referrer who enacts the “identify” skill: when learning new content, it observes the task, reflects on its previous knowledge and chooses which uncertainty or missing information to pursue.
  2. The second character is the detective, reflecting the “guess” skill: it formulates educated guesses and makes predictions about the missing information to pursue.
  3. The third character is the explorer reflecting the “seek” skill: it pursues uncertainty by asking the relevant questions or by exploring the relevant resources.
  4. The final character reflects the “assess” skill: it evaluates whether the inquiry resulted in learning progress with respect to the initial uncertainty.

The training paradigm is divided into two steps.

Figure 9 : The pedagogical goals and content of the four videos (first step of training)

Metacognitive skills they saw during the videos. For this we create a web-based platform we name “Kids meta-think”, where they are prompted to use these skills appropriately during a reading-comprehension task, using the help of conversational agents that have the same appearance and roles as the 2D characters presented during the videos.

The interaction with the agents is designed to be like the following: once participants finish reading a text, the first referee appears (representing the IDENTIFY skill) and guides them into identifying and formulating a knowledge gap. After validating this step, the detective agent appears to guide them into formulating an educated hypothesis about it (GUESS skill). Then the explorer appears (SEEK skill) to help formulate the appropriate divergent question. And finally, once the divergent answer submitted, we display a set of 3 pieces of information that may or may not contain the answer to the question. The second referee thus appears (ASSESS skill) to lead children into reflecting on these pieces of information and decide whether or not they bring them closer to closing their knowledge gap. The agents had predefined scripted behaviors and did not give any feedback regarding children’s inputs.

See Figure 10 for a concrete example of the agents’ dialogue for a given text.

Figure 3: Examples of the four agents’ utterances for a given text

Technical implementation of “Kids meta-think”

The behaviors of the four agents in terms of selection of the appropriate prompts as well as proposing a list of options (if requested) are entirely predefined and hand-scripted. Indeed, the agents are connected to a database containing the text resources and, for each one, a list of corresponding prompts relating to the four metacognitive skills, i.e., the utterances for the conversational agents for each skill. These utterances consist of sentences to remind the definition of the skill, its importance and how to use it. Each text is also associated with a list of 3 propositions for each skill. All of these resources have been previously hand-generated by the research team and validated by two teachers as to their pedagogical relevance.

During the interaction, the agent’s automaton composes the dialogue utterances in order to include the appropriate cuing strategies (see Fig. 11). We changed the utterances between the texts to avoid repetition in the different agents’ dialogue. This implementation has no natural language processing or generative artificial intelligence methods to manipulate the behaviors for the different agents. All data entered by participants was saved in a local database and was only evaluated post-experimentation, during the data

Figure 4: Screenshot of the ”Kids meta-think” platform during the training, given one text

Preliminary Results

We conduct a pilot study to test our training with 15 primary school students (Abdelghani et al., 2023b). Our evaluation is based on assessing the accessibility of our content and the pre-post intervention effects on the participants’ 1) metacognitive efficiency– which is a key facilitator for curiosity, 2) ability to ask curious questions about a task at hand and 3) their perception of the value of curiosity.

First, regarding the performance during training, the assessments of understanding of the videos were positive revealing a good accessibility of the first part of training (successes over 75%). In addition, during the training with “Kids meta-think”, we observed that all children succeeded to complete at least four cycles during the training (on the 8 proposed) with a mean percentage of correct cycles completed over 60%. For a reminder, a complete cycle is one where children identify an uncertainty, generate a hypothesis for it, ask the relevant question about it and then decide whether or not they could find the answer they’re looking for within a given list of propositions.

Second, regarding the effect of training, we observed a significant enhancement in children’s metacognitive efficiency as well as an increased curiosity-based behavior on question-asking task. This means children became more accurate in judging their own level of knowledge and then more behaviorally curious. Importantly, the metacognitive enhancement was mediated by performances obtained in the first part of training (declarative knowledge). While the second part that helps gain procedural knowledge these concepts, affects their ability to ask curiosity-driven questions. Following Anderson and Krathwohl’s taxonomy of learning [13], these observations can be rather expected since metacognitive sensitivity require declarative knowledge about what one does and does not know while asking curious questions requires procedural knowledge (i.e. being aware of a missing information and then formulating a relevant question to reach it).

However, we saw no enhancement in participants’ perception of curiosity with this first sample. This observation can be explained by the nature (i.e. digital interaction) and the short duration of the intervention (8 days). It is indeed well documented in social attitude literature that reliable and sustainable attitudinal changes often require longer interventions and/or with more realistic social interaction with teachers or between children.

Our pilot experiment successfully showed a proof-of-concept of the efficiency of the IGSA paradigm in training metacognition, as well as positive effects on curiosity-driven learning

Future works

For future directions (Years 2 and 3), thanks to the setting up of a Léa-Ifé network (application in progress) involving 10 classes in Bordeaux and the surrounding area, we will assess the effectiveness of the "IGSA framework" on a larger scale, following the methodological rules of a Randomized and Controlled Trial (RCT); this approach is expected in the psychoeducational sciences, and will allow the child to have more agency to create his or her own semantic incentives to curiosity. This design will enable us to study whether an increased agency might activate creative ideation associated with curiosity states involved in learning (PhD’s work program of C. Desvaux).

To tackle the issue of key role of curiosity in learning in the late adulthood, we also plan to conduct a research program addressing cognitive training needs with aging (Years 2 and 3). Indeed, as an alternative of pharmacological treatments, many researchers have mobilized their workforce in the design of computerized- cognitive training or rehabilitative programs for older adults and for Mild Cognitive Impairment patients. However, as in the education area, the training field is face to the central challenge of managing the diversity in response to cognitive training, which encompasses both inter-individual and intra-individual variabilities: some individuals positively respond to the trainings, others respond little or sometimes, or even, not at all).  

To meet the various needs of cognitive training with aging, we will explore adaptive methods of customizing the training program to each individual with the purpose of supporting curiosity-based learning by maximizing both their motivational/curiosity states and their training progress [1]. Especially, we plan to apply some principles from intelligent tutoring systems (ITS) to the field of attention training. Thanks to our collaboration with D. Bavelier, we have already developed automatic curriculum learning algorithms such as those developed in the KidLearn project in Flowers team [3, 5], which allow to customize the learner's path according to his/her progress and thus optimize his/her learning trajectory while stimulating his/her motivation by the progress made. The next steps will be to verify the correct behavior of our ITS, then to carry out an RCT study (adaptive vs. linear training) on young and elderly adults, and to investigate 1) the efficiency of the training during its execution and 2) its effects on cognitive tasks requiring attention (work program of PhD of M. Adolph and of post-doctoral position of M. Pèch in collaboration with Onepoint company).

Joint Publications

Joint communications

Other cited references

[1]        Oudeyer, P-Y., Smith, L. (2016) How Evolution May Work Through Curiosity-Driven Developmental Process, Topics in Cognitive Science 1–11.

[2]        Gottlieb, J.,Oudeyer, P-Y. (2018) Towards a neuroscience of active sampling and curiosity, Nature Reviews Neuroscience, 19(758–770).

[3]        Clement, B., Roy, D., Oudeyer, P-Y. *, Lopes, M. * (2015) Multi-Armed Bandits for Intelligent Tutoring Systems, Journal of Educational Data Mining (JEDM), Vol 7, No 2.

[4]        Ceha, J., Chhibber, N., Goh, J., McDonald, C., Oudeyer, P. Y., Kulić, D., & Law, E. (2019, April). Expression of Curiosity in Social Robots: Design, Perception, and Effects on Behaviour. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (p. 406). ACM

[5]        Delmas, A., Clément, B., Oudeyer, P-Y., Sauzéon, H. (2018) Fostering Health Education With a Serious Game in Children With Asthma: Pilot Studies for Assessing Learning Efficacy and Automatized Learning Personalization, Frontiers in Education, 3:99. doi: 10.3389/feduc.2018.00099

[6]        Sauzéon, H., N'Kaoua, B., Arvind Pala, P., Taillade, M., & Guitton, P. (2016). Age and active navigation effects on episodic memory: A virtual reality study. British Journal of Psychology, 107(1), 72-94.

[7]        Meade, M. E., Meade, J. G., Sauzeon, H., & Fernandes, M. A. (2019). Active Navigation in Virtual Environments Benefits Spatial Memory in Older Adults. Brain sciences, 9(3), 47.

[8]        Lotte F., Jeunet C., Mladenovic J., N’Kaoua B., Pillette L., (2018) « A BCI challenge for the signal processing community: considering the human in the loop », IET Book ‘Signal Processing and Machine Learning for Brain-Machine Interfaces’, Eds Tanaka & Arvaneh

[9]        Lotte F, Bougrain L, Cichocki A, Clerc M, Congedo M, Rakotomamonjy A, and Yger F, (2018) "A Review of Classification Algorithms for EEG-based Brain-Computer Interfaces: A 10-year Update", Journal of Neural Engineering, vol. 15, no. 3

[10]        Oudeyer, P. Y., Gottlieb, J., & Lopes, M. (2016). Intrinsic motivation, curiosity, and learning: Theory and applications in educational technologies. In Progress in brain research (Vol. 229, pp. 257-284). Elsevier.

[11]        Carstensen, Laura L.; Isaacowitz, Derek M.; Charles, Susan T. (1999). "Taking time seriously: A theory of socioemotional selectivity". American Psychologist. 54 (3): 165–8

[12]        Murayama, K., FitzGibbon, L., & Sakaki, M. (2019). Process account of curiosity and interest: A reward-learning perspective. Educational Psychology Review, 31, 875-895.

[13] Wilson, L. O. (2016). Anderson and Krathwohl–Bloom’s taxonomy revised. Understanding the new version of Bloom's taxonomy. The Second Principle, 1(1), 1-8.