Machine Learning and Human Learning
Machine learning is a powerful tool for understanding and predicting patterns in data, as evidenced
by DALL-E-2 and most recently ChatGTP-3. Despite these amazing feats, I’m of the opinion that they
are not well-suited for understanding human learning.
The primary issue with machine learning is that they are essentially black boxes where we don’t know why they’re making the decisions they make. This makes them hard to interpret and hard to improve upon unless you drastically change the underlying structure of the algorithm. For example, if a machine learning algorithm consistently makes an incorrect prediction, we would have no idea why. On the other hand, simple psychological models where we understand what each parameter corresponds to, it becomes much easier to understand why it makes the predictions it makes and further, if there are deviations in the prediction it’s relatively easy to iterate and improve upon the model.
Another large issue is data. Machine learning is data hungry. Hungrier than the cookie monster is for cookies. This is why image and text generation have been noticeably exceptional feats in machine learning, there’s heaps upon heaps of data. Humans make a lot of images on the internet (~7 billion per day in the past decade) and we’ve produced a lot of text in our entire history, most of which has been uploaded to the internet. This allowed for astounding training set sizes. DALL-E-2 was trained on about 400 million images. ChatGTP-3 was trained on 300 billion words. That’s a lot of data, and no analogous dataset exists in educational contexts and education data is hierarchical rather than discrete, making it more complex.
Further, the human learning system isn’t a monolith. Multiple systems exist within the mind and these systems likely have different rules that they play by. For example, learning math uses a different system than, say learning history or literature. Sure, both require the memorization of information, however, how that information is learned, processed, and represented is going to be different in each system. I think it is very unlikely that a machine learning algorithm used to predict how well someone learns math will also work to predict how well someone will learn history and it will not be obvious why. To make it even more complicated, multiple systems are used to learn tasks or information. For example, for motor tasks, like the piano, there are at least two systems involved in learning: the implicit motor system and the explicit declarative system. Thus, expecting the machine learning algorithm to either organize itself in a similar fashion or organize itself in such a way that its outputs are the same is a high bar to set that won’t be reached in the near future. This is also a problem for psychological models. Learning systems need to be represented in different ways, however, with psychological models, it is much easier to separate the contributions of each system to the performance output and model each individually within a larger system.
Overall, there are several advantages that I think psychological models will have over machine learning when trying to predict and enhance human learning. Since the space lacks large amounts of data and each learner can only generate so much, this disadvantages machine learning. In future posts, I’ll discuss how multiple learning systems interact in motor learning and the advantages of using data driven approaches to learning.