Adaptive learning software [DRAFT]
What would good adaptive learning software look like?
Cognitive load
Cognitive load is a well-supported observation in educational psychology that humans can only hold a few items (ideas, facts) in working memory at the same time. If we are presented with a large collection of new facts and ideas, and challenged with a problem requiring simultaneous manipulation of all these items, we perform very badly and learn little. Over time, however, we integrate concepts into schema, networks of interconnected concepts, which reside in long-term memory. These schema can then be accessed when needed to solve problems.
An analogy might be starting a new job. The morning of your first day is often stressful and chaotic: you need to think about what time to wake up, catch an unfamiliar bus at the right stop at the right time to the right place, get into the building, figure out where your new boss is, make sure you don’t do anything stupid in front of your new boss, find your desk, learn an unfamiliar computer system, figure out where to go for lunch…you will undoubtedly make many mistakes and be exhausted by the end of the day, from the effort of juggling so many new ideas and problems at once. One year into the job, however, you get from bed to desk on autopilot, and effortlessly handle routine problems associated with your work. The behaviours and conceptual chains which once had to be generated through manipulation of ideas in working memory have now been integrated into schemas stored in long term memory. (For a good discussion of cognitive load, see Ton de Jong: Cognitive load theory, educational research, and instructional design: some food for thought. Instructional science, vol 38 2009, pp105-134).
Good adaptive learning software would be conscious of cognitive load in four ways. Firstly, it would introduce new ideas hierarchically. This reduces cognitive load by taking maximal advantage of existing schemas. Secondly, it would introduce new ideas incrementally. New links in the semantic chains from which schemas are formed would be added one or two at a time. This reduces cognitive load by limiting the amount which has to be held in working memory at any given time. Thirdly, it would be parsimonious in the presentation of information, only showing the student the bare minimum required to comprehend the next conceptual step being developed. This again reduces cognitive load, by eliminating the need to consider irrelevant information, which needlessly occupies working memory. Finally, it would use Bayesian inference to track the existence and strength of schemas. By back-chaining through a hierarchical model of interrelated concepts which roughly approximates the student’s internal map, the software can infer areas of strength and weakness and direct learning accordingly.
Concepts linked in a semantic network
All teachers use an implicit or explicit network model of what they are trying to teach. This helps them to decide what concepts to introduce or withhold, in what order to introduce them, etc. To a greater or lesser extent, teachers also usually attempt to maintain at least a rudimentary understanding of the students' own internal models, and update this understanding as learning progresses. This allows teachers to, for example, notice when a student has missed an important piece of basic knowledge and correct for this.
Effective adaptive learning software must also have such a model, for exactly the same reasons. By offloading the work of presenting information and tasks and inferring student understanding onto software, having such a model supports and enhances the teacher’s ability to present information and problems at the right times, and to track student’s progress. For example, software which maintains a conceptual model for mathematics might use infer that a certain student attempting to learn quadratic expansion does not yet quite understand the nature of exponents. It could then present additional problems and material related to exponents to assist the student. If the student is still struggling and teacher intervention is required, the software could save the teacher time and effort by directing their attention towards the student’s problems with exponents.