Griffiths, T. L., Chater, N., Kemp, C., Perfors, A., & Tenenbaum, J. B. (2010). Probabilistic models of cognition: exploring representations and inductive biases. Trends In Cognitive Sciences, 14(8), 357–364. doi:10.1016/j.tics.2010.05.004

Summary

This article is one in a special issue and primarily discusses the differences between probabilistic models of cognition and connectionist models. The main claims of Griffiths et al. are that:

  1. Connectionist models are hard to interpret (what do the specific weights of each node mean?), even if they can accurately model behavior
  2. It is difficult, if not impossible, to specify inductive biases in connectionist models, whereas this is straightforward in Bayesian models
  3. It is difficult to specify structured representations in connectionist models
  4. It is difficult to use connectionist models to evaluate different hypotheses for representations, whereas Bayesian models can account for any type of representation, and are therefore a good tool for comparing and contrasting between different hypotheses
  5. All models (including connectionist models) build in hypothesis spaces, but probabilistic models make those spaces explicit

Importantly, Griffiths et al. do not claim:

  1. That Bayesian models should be evaluated against other types of models, but rather that Bayesian/probabilistic models provide a framework with which to work. Within that framework, different models can be represented and tested.
  2. That human cognition explicitly uses probabilistic inference, or that it is “optimal”. Rather, using rational analysis and finding the optimal solution the problem, and then comparing this to how humans behave, is a good methodology for studying cognition.

Methods

n/a

Algorithm

n/a

Takeaways

Probabilistic modeling is a powerful framework for studying cognition. It is not inherently opposed to connectionist models – other models (like connectionist ones) are actually a subset of the total models that probabilistic models can capture. Thus, it makes sense to work within the broader framework which gives us more flexibility for capturing and comparing a wide range of hypotheses for how behavior works.