To generalise, it seems that much e-learning related work taking place within educational institutions focuses on two key areas
- Learning Management System administration
- Authoring tool content generation
How do these areas relate to learning?
An LMS basically does the following:
- Presents content to learners in accordance with syllabi
- Tracks learners’ (instrumental) engagement with content
- Provides channels for learners to respond to content
- Tracks assessment of learners’ responses to content
- Provides (trackable) channels for learners to communicate with teaching staff and with each other
Authoring tools generate content for an LMS, and the content basically does the following:
- States declarative knowledge about subject matter to be learned
- Provides opportunities for automated practising of procedural tasks of a basically combinatorial character which are related to subject matter to be learned
Combinatorial learning is acquisition of procedural knowledge, where the procedures involved are easily and effectively reducible to combinatorial processes (such as selecting, matching, ordering, and recalling knowledge items from a collection of such items). Effective reducibility depends on either one of two conditions
- The collection of items is small
- The set of combinations that need to be known is small
To illustrate how combinatorial learning differs from generalised procedural learning, if arithmetic was learned through combinations, then answering a question like ‘3+8 = ?’ would be achieved by giving the answer ‘1’, then ‘2’, then ‘3’ etc… until reaching 11.
E-learning using authoring tool generated content with an LMS does not of course need to be as absurdly reductive as this. Combinatorial learning tasks can be automatically monitored such that excessively reductive attempts to complete them can prompt redirection to declarative knowledge content that is concerned with less reductive techniques for task completion and/or to more appropriately learner ability matched combinatorial tasks that would hopefully elicit less reductive responses.
This strategy to avoid reliance on reductive techniques is a significant aspect of what teaching involves. An effective teacher can apply a set of heuristics for detecting overly reductive task solving techniques and matching them with appropriate task redirections and declarative knowledge interventions. Teachers’ heuristics may well be in a large part known only implicitly, and so not readily translatable into automated procedures. This translation process is of many orders of magnitude greater in complexity than the processes of developing authoring tool content and configuring and maintaining LMS operations.
There is a concerning extent to which e-learning practice seems to be ducking the difficult challenge of formally attempting to explicate teachers’ heuristics into theoretical models and is instead focused on assisting in redirection of teachers’ professional efforts into generating content using authoring tools and engaging with LMS management. This approach may be based on an expectation that teachers’ heuristic expertise will, via a sort of aggregated cognitive osmosis, become ingrained in patterns of LMS recorded data from which said expertise could in principle later be statistically extracted. This has much in common with the way that social media platforms create value by accumulating and organising data about their users that can be used to predict their behaviour.
Recent AI research strongly tends to favour data driven approaches over model driven approaches, and recent research has led to practical improvements that appear to justify these approaches. In the case of education though there may be reasons to doubt whether this approach is the best one to take.
Data driven AI has a mimetic character. Recent notable AI achievements have tended concern things like natural language processing and object recognition in images; things that humans learn to do largely implicitly (there have been other, less obviously implicit, achievements of course, such as beating human masters at Go) but which have been difficult to model explicitly/formally.
If data driven AI teaching systems are generated via mimesis of human teaching, then it is presumably worth asking what sort of human teaching is being mimicked and whether this is the sort of teaching that it ultimately makes good sense to want to mimic.
Three principal modes of teaching can be identified- procedural, declarative and discursive.
Procedural teaching has a very long tradition- it goes back to apprentice craftsmen learned their crafts from their masters by observation, imitation, adoption of heuristics and through instrumental experimentation.
Declarative teaching can be nicely summed up by this (edited) extract from Hard Times.
“Girl number twenty,” said Mr. Gradgrind, squarely pointing with his square forefinger, “I don’t know that girl. Who is that girl?”
[her name is established to be Cecilia Jupe]
… “Cecilia Jupe. Let me see. What is your father?”
“He belongs to the horse-riding, if you please, sir.”
Mr. Gradgrind frowned, and waved off the objectionable calling with his hand.
“We don’t want to know anything about that, here. You mustn’t tell us about that, here. Your father breaks horses, don’t he?”
“If you please, sir, when they can get any to break, they do break horses in the ring, sir.”
“You mustn’t tell us about the ring, here. Very well, then. Describe your father as a horsebreaker. He doctors sick horses, I dare say?” “Oh yes, sir.”
“Very well, then. He is a veterinary surgeon, a farrier, and horse-breaker. Give me your definition of a horse.”
[Cecilia Jupe thrown into the greatest alarm by this demand]
“Girl number twenty unable to define a horse!” said Mr. Gradgrind, for the general behoof of all the little pitchers. “Girl number twenty possessed of no facts, in reference to one of the commonest of animals! Some boy’s definition of a horse. Bitzer, yours.”
“Bitzer,” said Thomas Gradgrind. “Your definition of a horse.”
“Quadruped. Graminivorous. Forty teeth, namely twenty-four grinders, four eye-teeth, and twelve incisive. Sheds coat in the spring; in marshy countries, sheds hoofs, too. Hoofs hard, but requiring to be shod with iron. Age known by marks in mouth.” Thus (and much more) Bitzer.
“Now girl number twenty,” said Mr. Gradgrind. “You know what a horse is.”
Blatant declarative teaching has become unfashionable in western education, but may be on its way back in since it has started to be seen by some policy makers as a major factor in the apparent successfulness of maths learning in many Asian countries (not necessarily a valid interpretation). Declarative learning has substantially given way to discursive learning; discursive in the sense of relating to discourse or modes of discourse, the emphasis being on higher order and critical thinking skills.
Some educators (notably David Didau in What if everything you knew about education was wrong?) argue that contemporary western educational methods are largely not genuinely discursive but rather remain declarative, except that the declarative learning occurs indirectly (and hence inefficiently)- learners in practice being coached and supported into superficially reproducing apparently discursive output which on closer examination consists of a heuristically aggregated mass of declarative knowledge that is concerned with discursive learning (learners do not actually learn to think critically but how to convincingly declare that they have thought critically). Didau argues that genuine discursive learning can only develop on a foundation of pre-existing declarative and procedural learning; I happen to agree with him.
Presuming that Didau is correct, then in my opinion the issue of what teaching should be mimicked boils down to whether procedural learning should emerge from declarative learning or vice versa.
The declarative to procedural pathway means first identifying the concrete items and objects that are to be learned about and then trying to understand how these fit together into structures. This is an analytical approach based on primarily acknowledging what needs to be learned and finding ways to accommodate how learners are able to learn to those needs. The primary tool for this pathway is instruction by experts. The experts could in principle be data driven AI teaching systems mimicking human teachers. These AI teaching systems should be able to teach much of what human teachers can, but it is not clear how they would necessarily improve on human teachers (other than in tireless availability). If the data that drove AI teachers was collected from one-to-one tuition (data on this being much less easily available than group teaching data however) then AI teachers could mimic such tutoring (and provide it at much lower cost than human tutors).
The procedural to declarative pathway means first experimenting with how processes of manipulating symbolic items and objects- which could well be facilitated through symbolic projection onto concrete objects that learners can actually (or virtually) manipulate rather than have presented to them declaratively- to generate structures, and then look for analogies between the structures generated and structures that exist in the world that need to be learned about. This is a synthetic approach based primarily on how learners are able to learn and finding ways to accommodate what needs to be learned to match these ways of learning. The primary tool for this pathway is the simulation/game and the community of its players- games providing the opportunities to manipulate, experiment and generate hypothetical structures (to play and to imagine) and communities of players providing the connections to real world structures through their embodied experiences of how these structures (the declarative knowledge about the world) do and do not relate to game generated structures.
The procedural to declarative pathway of course incorporates AI in the design of learning games, but a diverse set of specific model based AI systems, matched with various different real world systems. Each specific model would be designed through explicit processes of consultation with experts in a particular subjects or applications of subjects as well as using data about how learners’ played such games.
What I am suggesting is that the procedural to declarative pathway provides learning potentials that go beyond what could be achieved by automating existing teaching methods, primarily in terms of helping to develop learners’ creativity and problem solving skills. Learners would not just be learning to reproduce existing knowledge but learning how to generate new knowledge, and even how to generate tools for the generation of knew knowledge (by writing their own games).
To some extent, the (LMS + authoring tool generated content = data = learning) approach recognises the potential for new knowledge to be generated through LMS enabling of user generated content, where learners are the users generating the content. Learner to learner sharing has scope for having the same sorts of uses as communities of gamers sharing games and gameplay experiences with each other, and generic user-generated-content models certainly have far greater overall flexibility, but such flexibility can simply end up as a channel for the distribution of pseudo-discursive declarative knowledge- especially if prioritising such knowledge continues to comprise a significant part of educational practice.