Probabilistic models

The world is a fundamentally noisy and variable place: few events must occur; our measurements are rarely perfectly accurate; and relations are almost never deterministic in nature. Instead, there is uncertainty and error of various types in all our experiences, as shown by just the slightest reflection on everyday life. Sometimes, caffeine helps me to be more alert, but not always. Sometimes, my dog barks at strangers, but not always. Nonetheless, cognitive systems (including people) must be able to learn and reason appropriately despite this ineliminable noise and uncertainty. And in addition to variability in our experiences, human behavior is itself noisy and uncertain; people do not (and often should not) act identically in seemingly identical situations or contexts. Computational models of human cognition must have some way to handle all of the noise, uncertainty, and variability; many do so with probabilities, as the probability calculus is a standard computational framework for capturing and working with noise and uncertainty, whether in the world or the reasoner. ¹ As one illustrative example, almost all theories of category judgments (such as “Is this a dog?”) are probabilistic in nature: they allow for uncertainty in both the world – the same observation might sometimes be a dog, sometimes a wolf – and in the human categorizer – the same observation can probabilistically yield one of several possible judgments.