[tel]

This article finally revealed to me the mathematical structure of curriculum optimization. It's a bias/variance tradeoff.

If you've done statistics, you know that B/V tradeoffs are more or less an unavoidable feature of optimization or learning. If you go in with less clear goals, you depend on learning on the fly to find the best solution which means your performance varies a lot depending on environmental conditions. If you go in with very clear goals, it's likely that people will cluster around them tightly, but if you're wrong with your bias, you'll be surely fucked.

How do you beat B/V tradeoffs? You put in more effort, more experience, better sharing of information. You select your biases very carefully such that they are less clear but cannot possibly be wrong. You constrain your variance such that it is less harmful to your task.

Finally, there's the idea of consistency and convergence rates. Consistency means that if any person spends enough time and effort they will eventually overcome both bias and variance and find the best solution. Rate of convergence is how much time and effort it takes to get reasonably close. Consistency and fast convergence are highly valuable properties, obviously, but both are easy to hurt and destroy, especially through biasing.