Anthropodidactic machine learning is using didactic materials developed for human students (textbooks, lectures and/or lecture notes, explanations, homeworks, exercises, games and other sorts of interactive edutainment) to train artificial intelligence. Examples of anthropodidactic learning include using language textbooks to train a machine translation model or using a flight simulator developed for pilot training to train an autopilot with reinforcement learning [Staudinger Jorgensen Margineantu 2018].
Because the education industry puts a lot of effort into curating and systematising knowledge in a way that can be useful for any learner, whether they run on carbon- or silicon-based hardware. For example:
- Exercise sets in mathematics, physics and language learning, to name a few fields, are explicitly designed to cover all important clusters/corner cases of the subject area - something that isn’t guaranteed in most datasets like logs, business records or text corpora.
- Exercise sets are also sorted by difficulty. This creates a useful curriculum [Soviany 2021] to follow when training a machine learning model.
- Educational software aims to give users immediate and precise feedback on their mistakes: delayed gratification, as it turns out, is hard for people and reinforcement learning algorithms [Gulwani et al 2017] alike.
See The Art of Problem Posing [Brown, Walter 2004] to learn more about… the art of problem posing.
Machine learning community is undoubtedly interested in taking lessons from human learning, efforts to do so bear the umbrella term of antropomorphic machine learning. The prime example is curiculum learning [Soviany 2021]: it was born with the observation that the order in which data is presented to human students is crucial for them achieving their learning goals, so perhaps it makes a difference for machines too.
However, examples of directly reusing learning aids developed for human students are hard to come by. A notable exception is Reinforcement Learning where decision-making agents are often trained on games initially intended for people. And while the claim that Atari games and Minecraft are educational material may be somewhat stretching the definition of education, interactive simulators first developed for people and later adapted for reinforcement learning include X-plane (used for training pilots) and Virtu-ALS [Liventsev et al 2021] (used for training nurses). Some antropodidactic work has also been done in natural language processing, training language models on children’s books [Hill et al 2016] and exercises for language learning [Mayhew et al 2020] from Duolingo
These examples, however, are exceptions rather than the rule and, in general, anthropodidactic programming remains criminally underexplored. A couple of research directions that seem extremely promising to me are:
- Training a language model on exercisebooks in subjects like mathematics and the sciences to achieve a system capable of problem solving in these fields.
- Using beginner-level programming tasks to develop program synthesis.
Didactic materials are a large class of useful data waiting for someone to turn them into a successful artificial intelligence system/product. Will it be you, %username%?
If you’re interested in working on this topic (or already are!) feel free to send me an email, let’s see if we can collaborate.
This work is taxpayer funded