Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Introaied nancy2019 luengo


Published on

Introduction to AI in education domain. Description of intelligent tutoring systems

Published in: Science
  • Be the first to comment

  • Be the first to like this

Introaied nancy2019 luengo

  1. 1. Document confidentiel – ne peut être reproduit ni diffusé sans l'accord préalable de Sorbonne Université. TEL and Artificial Intelligence bring to Education 1
  2. 2. QUI JE SUIS ? EDTECH masterhttp://sciences.sorbonne- ANDROIDE
  3. 3. 3 LIP6
  4. 4. 4 LIP6
  5. 5. 5 LIP6, 22 teams
  6. 6. 6 7 researchers 6 PhD students + 5 finished in 2019 2 postdoctoral researchers 1 contractual engineer Several masters students => Currently we have two postdoctoral open LIP6 MOCAH team
  7. 7. 7 MOCAH: Current research Cognitive diagnostic and learner assessment ▪ In mathematics ▪ In physics ▪ In medicine Adaptation, personalization and feedback ▪ Context awareness ▪ Differentiated feedback to learners / teachers / designers / institutions ▪ Highly interactive systems ▪ Massively multi user learning systems ▪ Ubiquitous environments ▪ Collaborative Learning games Authoring tools ▪ Meta-design / Co-design ▪ For end-users (teachers) Traces analysis ▪ Group detection, peer recommendation ▪ Knowledge pattern extraction (from forums, interactions, etc.) ▪ Behavior pattern extraction Merging symbolic and numerical approaches Applied and Multidisciplinary research
  8. 8. 8 At the French level
  9. 9. 9 International level
  10. 10. 10 But also => EDM 2021 will be in PARIS, Sorbonne Université
  11. 11. 11 Today This morning introduction to AIED domain • Historical point of view to introduce the classical models Afternoon an example that I developped during 11 years • Intelligent tutoring system in ortopedic surgery
  13. 13. 13 Some questions What is AIED? What can it bring? How? How to connect AI to teaching and learning? Risk: adapting to what is liked, not what is learned Making an impact at the right level: teachers & students level Risk: more efficient administration, not more efficient learning Ethics: what acceptable uses? Risk: new ideas once the data is there
  14. 14. 14 Goals today Titre de la présentation Not replacing the teachers: harness the strengths of AI to empower them Learning that is more: Personalized Flexible Inclusive Engaging Respond to what is being learned, how it’s being learned, what is felt… The dilemma: Decades of research: great ideas are there (Cheap) technological devices are there Practical applications… not so much R. Luckin, W. Holmes, M. Griffiths & L. B. Forcier Intelligence Unleashed – an Argument for AI in Education”, 2016
  15. 15. 15 What is AIED? Titre de la présentation 30 years of research AI + learning sciences (education, psychology, neuroscience, linguistics, sociology) to promote adaptive learning environments Goal: opening the black box of learning Methods: Using theories Building models (knowledge about the world) Process with algorithm Testing Iterating
  16. 16. 16 What models? Titre de la présentation Domain model: what to teach, to learn Learner model: students’ status (frequent update) Achievements, skills, difficulties Emotional state Engagement (time on task) Other models => metacognitive, emotional, social… Pedagogical model: how to teach Productive failures (mistakes are ok) Appropriate feedback (which hints, when…?) Testing knowledge (how?)
  17. 17. 17 Classical AIED models => first period (ITS) Titre de la présentation Domain model Pedagogical model Learner model Algorithms: decision making process Learning content individualized xAdaptive learning environment
  18. 18. 18Titre de la présentation Domain model Pedagogical model Learner model Algorithms: decision making process Learning content individualized xAdaptive learning environment Logs Students’ actions, Affect, Speech… Algorithms: data analysis Feedback Updates Updates Open Learner Model Classical AIED model => now R. Luckin, W. Holmes, M. Griffiths & L. B. Forcier Intelligence Unleashed – an Argument for AI in Education”, 2016
  19. 19. 19 History of AIED = ITS Precursors examples SCHOLAR (Carbonell, 1970) This program attempted to engage the student in a mixed initiative dialogue on South American geography. The program and student communicated through a sequence of natural language questions and answers. The tutor could both ask and answer questions and keep track of the ongoing dialogue structure. SOPHIE (a SOPHisticated Instructional Environment) (Brown, Burton & deKleer, 1982) It was a tutor for electronic troubleshooting that used its expert system to evaluate the measurements students were making in troubleshooting a circuit. GUIDON (Clancey 1982) an ITS for teaching medical diagnostic problem solving first attempt to adapt a pre-existing expert system into an intelligent tutor. It attempts to transfer expertise to the students exclusively through case dialogues where a sick patient (the 'case') is described to the student in general terms.
  20. 20. 20 History of ITS, before 1980’s Scholar SCHOLAR was a pioneering effort in the development of computer tutors capable of handling unanticipated student questions and of generating instructional material in varying levels of detail, depending on the context of the dialogue. It was a mixed-initiative ITS: both the system and the student could initiate conversation by asking questions. Both the program's output and the student's inputs were English sentences.
  21. 21. 21 Tuteurs, Scholar > approx what is the area of Brazil ? S : 2 500 000 square miles > wrong. Please indicate if the following statement is correct: the area of Paraguay is approx 47 432 square miles S : what is the area of Brazil > the area of brazil is approx 3 600 000 square miles. Now answer the question you didn't answer before. S : false > rephrase your statement. I can't understand it S: incorrect >very good. The area of Paraguay is approx 157 000 square miles. The language of Brazil is French. Right or wrong ?
  22. 22. 22 Tuteurs, Sholar > The eastern and central plain of Argentina is... S : Chaco > No. The northern plain of Argentina is the Chaco. The eastern and central plain of Argentina is the Pampas. S : is the Chaco in Chile? > No. The Chaco is in Argentina, Paraguay and Bolivia, not in Chile. S : are the Andes a region of Chile > Well, you are almost right. More specifically, a region of Chile is the Southern Andes.
  24. 24. 24 Domain model Source: Wenger, Etienne (1987) Artificial Intelligence and Tutoring Systems: Computational and Cognitive Approaches to the Communication of Knowledge Domain to be conveyed to the student, i.e. the knowledge of the experts or how experts performs the domain. dynamic model of the domain knowledge and a set of rules by which the system can "reason.“ Roots in expert systems research (such as medical diagnostic or electronic troubleshooting systems) in some ITS it have the ability to generate multiple correct sets of solutions, rather than a single idealized expert solution. => contains the concepts, rules, and problem-solving strategies of the domain to be learned. It can fulfill several roles: as a source of expert knowledge, a standard for evaluating the student’s performance or for detecting errors, etc.“
  25. 25. 25 Scholar the expert knowledge module : geography of South America. Represented in a semantic network whose nodes instantiated geographical objects and concepts. Statements like 'Tell me more about Brazil' just invoked a retrieval of facts stored in the semantic network. The real power of this representation schema comes by recognizing that it is possible to answer questions for which answers are not stored. it is not necessary to store in the semantic network that 'Lima is in South America‘. the program must know about the attributes concerned, e.g. 'location' and 'capital', and in particular, that if x is capital of v and y is located in z then x is in z: this is a rule of inference. Carbonell, J.R, 1970, AI in CAI: An AI Approach to CAI, IEEE Transactions on ManMachine Systems, V. 11, pp, 190
  26. 26. 26 Domain model, another example Cognitive tutors : lessons learned Anderson, J. R.; Corbett, A. T.; Koedinger, K. R. & Pelletier, R. 1995
  27. 27. 27 Domain model, several approches Several theories A Comparative Analysis of Cognitive Tutoring and Constraint-Based Modeling, Mitrovic, Koedinger and Martin, User Modeling 2003, pp 313-322 Constraint base Olhson’ errors theory Model tracing Anderson ACT theory
  28. 28. 28 Domain model, limits KnowIedge elicitation and codification can be a very time- consuming task, especially for a complex domain with an enormous amount of knowledge and interrelationships of that knowledge.  investigating how to encode knowledge and how to represent it in an Intelligent system remains the central issue of creating an expert knowledge module. => the problem of ill-defined domains remains also a research question
  29. 29. 29 Domain model Hybrid aproaches Semantic (rules, semantic networks) + numerical approaches • combine the advantages of different approaches in order to overcome their limitations. • different approaches can be better suited for different parts of the same task so as to offer common or complementary tutoring services.
  30. 30. 30 Domain model Hybrid aproaches in CanadarmTutor (a robotic arm deployed on the international space station which has seven degrees of freedom) there are 1. There are not a good solution (or hard to ellicitate) 2. no clear strategy for arm manipulation as it moves from the original configuration to the targeted configuration. => Several kind of knowledge : procedural, spacial, declarative…
  31. 31. 31 Domain model Hybrid aproaches 3 kinds of models 1. an expert system (Kabanza et al. 2005) 2. a cognitive model (Fournier-Vigier et al. 2008) => model tracing 3. the partial task model approach (Fournier- Vigier et al. 2009, 2012) => Mining sequential rules Canadarm Tutor provide assistance for different parts of the arm manipulation task. The result is tutoring services which greatly exceed what was possible to offer with each individual approach, for this domain (Fournier-Vigier et al. 2009). Fournier-Viger, P., Faghihi, U., Nkambou, R., Mephu Nguifo, E.. CMRules: Mining sequential rules common to several sequences. In Journal Know.-Based Syst. Vol. 25(1), (2012) 63-76
  32. 32. 32 Domain model Hybrid aproaches • Itemset • Sequence • Algorithm CMRule 2 MOOC EIAH Fournier-Viger, P., Faghihi, U., Nkambou, R., Mephu Nguifo, E.. CMRules: Mining sequential rules common to several sequences. In Journal Know.-Based Syst. Vol. 25(1), (2012) 63-76 {x y} => {z} means that at each occurrence of items x and y, we observe an occurrence of item z (given support and confidence) Figure from Fournier-Vigier 2012, p. 64
  33. 33. 33 Domain model Hybrid aproaches “Canadarm Tutor provide assistance for different parts of the arm manipulation task. The result is tutoring services which greatly exceed what was possible to offer with each individual approach” Fournier-Viger, P., Faghihi, U., Nkambou, R., Mephu Nguifo, E.. CMRules: Mining sequential rules common to several sequences. In Journal Know.-Based Syst. Vol. 25(1), (2012) 63-76
  35. 35. 35 Student model and diagnosis Dynamic representation of the emerging knowledge and skill of the students and it describes how to reason about their knowledge COMPUTER MODEL OF THE STUDENT ≠ STUDENT COMPUTER MODEL Data about student How the student reason… DIAGNOSIS = Process of inferring the current state of the knowledge
  36. 36. 36 Student Model Source: Wenger, Etienne (1987) Artificial Intelligence and Tutoring Systems: Computational and Cognitive Approaches to the Communication of Knowledge According to Wenger, student models have three tasks : • They must gather data from and about the learner. This data can be explicit -- asking the student to solve specific problems -- or implicit -- tracking the students navigation and other interactions and comparing them to information about similar learner responses. • They must use that data to create a representation of the student's knowledge and learning process. The system then uses this model to predict what type of response the student will make in subsequent situations, compares that prediction to the students' actual response, and uses that information to refine the model of the student. • The student model must account for the data by performing some type of diagnosis, both of the state of the student's knowledge and in terms of selecting optimal pedagogical strategies for presenting subsequent domain information to the student.
  37. 37. 37 Student Model Titre de la présentation ALGEBRA TUTOR Correct rules + « bug » rules
  38. 38. 38 Student Models errors models examples with aproaches enumerative reconstructive generative enumerative ACTP, BUGGY Extensible list of errors DEBUGGY, LMS : combination of enumerated errors that allow the reconstruction of observed errors MEMO-II : List of enumerated errors with a link to enumerated misconceptions reconstructive PROUST: conception reconstructive from intentions using a library of plans error ACM, PIXIE, ADVISOR : errors reconstructed from a neutral language with primitive Young & O’Shea (1981) : Incorrect procedures that are reconstructed from the explanation of the nature of the errors generative Bonar & Soloway (1985) : abstract library errors that have been generated to explain the origin of the observed errors REPAIR : Errors generated by repetition of processing impasses REPAIR/STEP, matz (1982) : Reduction of the occurrence of errors by “misslearning” Source Wenger 1987, p. 348
  40. 40. 40 Numérical approche example Knowledge tracing Knowledge tracing has become the dominant method of modeling student knowledge. Knwoledge tracing is a predictive model See : Corbett and Anderson 1995
  41. 41. 41 Student model Knowledge tracing approach Corbett and Anderson 1995
  42. 42. 42 Knowledge tracing in others words At each successive opportunity to apply a skill, KT updates its estimated probability that the student knows the skill, based on the skill-specific learning and performance parameters and the observed student performance (evidence). Reye showed that KT is special case of a DBN which assumes parameters do not change across time slices. => the conditional independence graph of KT can be drawn as the following Figure. Corbett and Anderson 1995
  43. 43. 43 Student model, from data Titre de la présentation ALGEBRA TUTOR
  44. 44. 44 Student model knwoledge represented Source : Woolf 2009
  46. 46. 46 Challenges for 2030 The student model We envision that by 2030 user models for students will be complex, not only representing what students know, do and have abilities for, but other factors too. For instance, user models will track when and how skills were learned and what pedagogies worked best for each learner. Moreover, user models will include information on the cultural preferences of learners, their personal interests, learning goals, and personal characteristics, to select the optimal mix of learning environments, pedagogy, visualizations, and contexts that maximize engagement, motivation and learning outcomes for each individual. When the learner is part of a group, the model will make the best compromise among the individuals who are part of the group. European network of excellence
  47. 47. 47 Challenges for 2030 the architecture Most likely, by 2030 user model servers will be readily available for education. Servers are similar to generic user models in that they are separate from the application and will not run as part of it. User modelling servers will be part of local area networks or wide area networks and serve more than one application instance at a time
  48. 48. 48 Challenges 1. sharing of learner models across learning systems In the long-term, this trend may lead to a more integrated and effective educational experience for students, across their life-time of learning. In the long-term, as the field gets better at developing, refining, and exploiting => sophisticated multi-dimensional models of learners, there is improved potential for tailoring each student’s learning experiences to their educational needs. M. C. Desmarais, R. S. J. d. Baker, A review of recent advances in learner and skill modeling in intelligent learning environments. User Model User-Adap Inter (2012) 22:9–38
  50. 50. 50 History of ITS 1980’s Architecture : Pedagogical model Source: Wenger, Etienne (1987), Designs and regulates instructional interactions with the student. Represent teaching strategies and includes methods for encoding reasoning about the feedback Teaching strategies : examples, analogies, …. learning is viewed as successive transitions between knowledge states, the purpose of teaching is accordingly to facilitate the student's traversal of the space of knowledge states." (p. 365)
  51. 51. 51 But… More systems : Inform the diagnosis to the human Propose a direct feedback for each diagnosis. If diagnosis=X them feedback=Y => Few systems propose an independent computer model which models the decision of a pedagogical feedback following a diagnosis.
  52. 52. 52 Feedback in Cognitive Tutors provide immediate feedback after each problem-solving step. The feedback uses rules and mal-rules of the student model Source : Anderson & all 1995 (Cognitive Tutors, Lessons Learned)
  53. 53. 53 Pedagogical model : feedback example Rule that gives the solutions
  54. 54. 54 Feedback in Cognitive Tutors
  55. 55. 55 Feedback in other kind of model Tracing tutors (Andes) Ref : VanLehn et all. 2005, The Andes physics Tutors, Lessons Learned
  56. 56. 56 Feedback in other kind of model Tracing tutors (Andes) Three kinds of feedbacks Flag feedback : red incorrect and green if it is correct In order to give immediate feedback, entries relevant to solving the problem and the solution point for the problem. In principle, this information could be provided by a human author instead of being the solution graph file needs to contain the set of nonequation generated by Andes”. Explications feedbacks (What's Wrong Help) In order to implement What's Wrong Help, Andes needs only three sources of knowledge: the knowledge base of error handlers, one solution point per problem, and one set of defined quantities per problem Hints to solve the problem (Next Step Help) it gives a sequence of hints intended to accelerate learning. First Implementation : The first version attempted to recognize the student’s plan for solving the problem and hint the next step in that plan. The Andes1 Bayesian net was used for both plan recognition and next step selection. But p is not consistent with expert human help. the new version of Next Step Help does not try to recognize the student’s plan, but instead it engages the student in a brief discussion that ensures the student is aware of the principle that is being used to solve the problem, then selects a step the student has not yet done, and hints that step => Feedback based in domain model. => Feedback based student model (the solution graph with errors)
  57. 57. 57 Feedback models based in Bayesian student models Given a Bayesian student model, the next issue is how to use the model to optimise the pedagogical actions of the intelligent tutor. Source : Mayo et al 2001. Optimising ITS Behaviour with Bayesian Networks and Decision Theory
  58. 58. 58 Feedback models based in Bayesian student models three general approaches Alternative Strategies Diagnostic strategies Decision-theoretic pedagogical Source : Mayo et al 2001. Optimising ITS Behaviour with Bayesian Networks and Decision Theory
  59. 59. 59 Feedback models based in Bayesian student models three general approaches Alternative Strategies : optionally take the posterior probabilities of the Bayesian network and use them as the input to some heuristic decision rule. ADELE (Ganeshan et. al., 2000). ADELE has a Bayesian network model of the domain knowledge, but it uses a heuristic based on focus-of-attention to select the node in the network about which to provide a hint. Andes and SQL tutors was used also heuristics decisions rules at different levels Source : Mayo et al 2001. Optimising ITS Behaviour with Bayesian Networks and Decision Theory
  60. 60. 60 Feedback models based in Bayesian student models three general approaches Diagnostic Strategies: The basic idea is to select actions whose outcomes are likely to maximise the posterior precision of some node in the network. Millán et. all domain is test question selection, and questions are selected to maximise the system’s certainty that the student has mastered the domain concepts. This strategy has limited applicability outside of diagnostic tests. Source : Mayo et al 2001. Optimising ITS Behaviour with Bayesian Networks and Decision Theory
  61. 61. 61 Feedback models based in Bayesian student models three general approaches Decision-Theoretic Pedagogical Strategies: select tutorial actions that maximise expected utility. While diagnosis is obviously an important component of expected utility maximisation, it is only a secondary component. The primary consideration of an expected utility calculation is the likely outcomes of the action, and their pedagogical utility in CAPIT the expected utility of an action (e.g. problem selection) depends on the likely outcomes of the action (e.g. how many errors are made). The impact of the feedback could be calculated with some type of factors In DTTutor, the action’s impact on many different factors related to the student (e.g. their morale, etc) has an influence on expected utility. Source : Mayo et all 2001. Optimising ITS Behaviour with Bayesian Networks and Decision Theory
  62. 62. 62 BN Extension : influence diagram or decision network RB Use the value information to decide. Inference of the utility based in a value information Three type of nodes Chance nodes : represent random variable, like in clasical BN Decision nodes : represent points where decisions maker has a choice of actions Utility nodes (or value nodes) : represent the utility function. Function that maps from states to real numbers
  63. 63. 63 Feedback, computer considerationsCategory Differentiation Modality Monomodale, multimodal Adaptation Non adaptive, adaptive (macro adaptive, micro-adaptive) Personalization Non-personalized, Personalized Independency of the model Strong coupled, weakly couple, independent Automatic Automatic, semi-automatic Intelligence Non intelligent Intelligent optimization strategy Linguistic Alternative Diagnostic Decision Theory Textual Oral
  64. 64. 64 Other numerical pedagogical model AI for Education - F. Bouchet
  65. 65. 65 Perspectives pedagogical model « A single teaching strategy was implemented within each tutor with the thought that this strategy was effective for all students. However, students learn at different rates and in different ways, and knowing which teaching strategy (…) is useful for which student would be helpful. This section suggest the need for multiple teaching strategies within a single tutor so that an appropiate strategy might be selected for a given student » Woolf (2009, p. 133)  Several statrategies  Choose the better at the right time => DATA NOT ONLY FROM STUDENT BUT ALSO FROM TEACHERS, TUTORS;;; AI for Education - F. Bouchet
  66. 66. 66 Perspectives Assesment and feedback […] assessment instrument can assist the teacher in giving good feedback to the student. However this takes time and developing automatic marking and feedback systems for formative assessment will assist with the scalability of this proven effective pedagogical practice. Timely feedback provides a motivating and important experience for learners and the technology can reduce the over- reliance of the teacher as the sole assessor. Source stellar Document
  67. 67. 67 CURRENT AIED… C AI for Education - F. Bouchet
  68. 68. 68 Current uses of AI for Education Personal tutors (ITS) Intelligent support for collaborative learning (CSCL) Intelligent virtual reality 68 AI for Education - F. Bouchet
  69. 69. 69 Personal tutors : intelligent tutoring systems Individualized tutoring: an ideal but cannot scale… until now Different pedagogical approaches: Scaffolding learning: between support and challenge Diagnosing procedural errors (BUGGY) Helping learners to be in control: self-regulated learning « learning how to learn » (MetaTutor) (Azevedo et al. 2013) Different methods to get there: Symbolical: requires experts uses models, ontologies… Numerical: requires data uses self-training algorithm (machine learning) 69 AI for Education - F. Bouchet Hard to build Hard to interpret
  70. 70. 70 Supporting colloborative learning Collaboration helps learning: Pair of students in online courses have higher learning outcomes than students learning alone Encourages reflexion Caring about the group increases engagement AI can help: To form efficient groups with complementary skills (Labarthe et al. 2016) Identify efficient collaborative patterns (to help students or teachers) Through pedagogical agents: tutor (AutoTutor), peer (learning by teaching (Betty’s Brain)(Biswas et al. 2012)) Intelligent moderation: detect off topic discussions 70 AI for Education - F. Bouchet
  71. 71. 71 Intelligent virtual reality VR can help with learning: Encouraging « what if » scenarios (simulation) Visiting historical places Enabling low-achieving students by shifting their self image (Crystal Island) AI can make virtual world « intelligent »: DIRE QUELQUE CHOSE SUR LES GESTES ET PERCEPTIONS Guiding learners to regulate their emotional status Encouraging collaboration between learners Applications: Against bullying (FearNot!) (Vannini et al. 2011) Peacekeeping scenarios (Traum et al. 2003) 71 AI for Education - F. Bouchet
  72. 72. 72 FUTURE AIED… C AI for Education - F. Bouchet
  73. 73. 73 Future uses of AI for Education ? Where is AI going? What are the challenges? Develop reliable indicators to track progress More and more capture devices (biological data, eyetracking, speech recognition…) Better understanding of the best teaching approaches and their context More data collected + sharing 73 AI for Education - F. Bouchet
  74. 74. 74 « A Renaissance in Assessment » Just-in-time assessments Today: LA can predict if a student will fail or drop-out Tomorrow: how motivation and engagement varies Tracking learning progress Today: is the answer right/wrong? Tomorrow: why? What type of mistakes? What emotional state? Stealth assessments: Today: short quizzes, final exam Tomorrow: assessing while learning is happening (e.g. through a collaborative project) 74 AI for Education - F. Bouchet (Hill and Barber, 2014)
  75. 75. 75 New insights from learning sciences AIED will be more interdisciplinary than ever Education neuroscience: Today: Uncertain rewards can improve learning (Howard-Jones et al., 2014) Tomorrow: calibrating rewards based on learner Psychology: Today: « growth mindset » in learners is more efficient than « fixed mindset » (static intelligence) (Dweck, 2010) Tomorrow: detecting student’s mindset and develop it 75 AI for Education - F. Bouchet
  76. 76. 76 Lifelong learning partners We learn more efficiently with another (Cole, 1996) Today: learner-companion help stimulate student learning in various ITS Tomorrow: one assistant companion Across app and devices In and beyond school Choosing the optimal resources when they are needed 76 AI for Education - F. Bouchet
  77. 77. 77 Tackling unsolved issues Achievement gap – a social issue: Help those who need it the most with one-on-one tutoring Making sure all benefits from AIED Developing teacher expertise: Reducing stress Freeing time to do what humans do best with students (automatic grading, resource recommendations…) 77 AI for Education - F. Bouchet
  78. 78. 78 What about ethics? Overall problem with data collection: For what? For whom? Who decides? Problem with algorithms: What happens when AI goes wrong? Who is responsible? Sharing is necessary for AIED to be successful Guaranteed anonymousness (privacy by design) AIED encourages human behavior changes & to establish relationships Should every mistake be reported to the teacher? Spy effect 78 AI for Education - F. Bouchet
  79. 79. 79 Conclusion AIED exists today, and will be more and more important in years to come It needs institutional support to spread: Involve learners, teachers, parents in the co-design of the next AIED systems to meet their needs Develop, evaluate… iterate Spread data standards Share data, but keeping the ethics in mind! 79AI for Education - F. Bouchet
  80. 80. 80 Thanks to… KIWI for the invitation to the summer school François Bouchet for the first version of this presentation “Intelligence Unleashed – an Argument for AI in Education” by R. Luckin, W. Holmes, M. Griffiths & L. B. Forcier – on which this presentation is loosely based on You for your attention! 80 AI for Education - F. Bouchet