14
$\begingroup$

In recent years, Bayesian models of cognition have been used - with considerable success - to explain human reasoning in a variety of inferential tasks (Chater, Tenenbaum, & Yuille, 2006). These models represent a "probabilistic approach" that seeks to derive the optimal solutions to inferential problems, and the success of these models is often interpreted as evidence that intuitive reasoning is fundamentally rational. Other successful research traditions, however, adopt a starkly different view of human rationality. Researchers interested in judgment and decision making, for example, often assume that intuitive inference is heuristic, error-prone, and subject to biases (Kahneman & Tversky, 1982; Gilovich, Griffin, & Kahneman, 2002).

How is it possible that the probabilistic and heuristic-and-biases approaches have both been successful while adopting fundamentally incompatible views of rationality? I suspect that many psychologists have thought about this question, but so far as I can tell, these approaches have only been directly compared on a few inferential tasks, and there are still fewer (published) attempts to derive any general answers to this question.

References

  • Chater, N., Tenenbaum, J.B., & Yuille, A. (2006) Probabilistic models of cognition: Conceptual foundations. Trends in Cognitive Sciences 10(7): 287-291. [pdf]

  • Gilovich, T., Griffin, D., & Kahneman, D. (Eds.) (2002) Heuristics and biases: The psychology of intuitive judgement. Cambridge Univ. Press.

  • Kahneman, D. & Tversky, A. (1982) The psychology of preferences. Scientific American 246(1): 160-173.

Improve this question
$\endgroup$

2 Answers 2

Reset to default
7
$\begingroup$

Your question is predicated on the assumption that Bayesian modeling has been successful in all domains. I think this is a stance that many (except hardened Bayesians) would disagree with. For instance, consider the classic Tversky & Shafir experiments on the violation of the sure thing principle:

What are popular rationalist responses to Tversky & Shafir?

The above approach can be reconciled with standard rationalism, but you are in fact building theories of heuristics-and-biases to do so; that is how Kahneman & Tversky arrived at their models. Similarly, you can build very awkward Bayesian models for these sort of tasks (other difficult tasks are ones that show an order effect) but it is hard to say how these models are different from heuristics. For a more general treatement:

What tasks does Bayesian decision-making model poorly?

However, in the domain where Bayesian modeling is a good fit, the reason it works well to reconcile with heuristics-and-biases is because it has the perfect mechanism for modeling bias: the initial prior distribution. In fact you can use the prior distribution in the bayesian approach to quantify bias.

To summarize, you can model heuristics-and-biases with the bayesian approach. For bayesians heuristics are how the task is modeled/represented and biases are the intial prior distributions

Improve this answer
$\endgroup$
3
  • $\begingroup$ Good response. I'd like to add that the brain feeds it's own activity back onto itself. So if you use a Bayesian model of learning, it's not simply collecting external information; the prior state of the system is being constantly factored into the new state along with external information. $\endgroup$
    – Preece
    Commented Jun 25, 2012 at 17:53
  • $\begingroup$ Thanks for the answer Artem! I tend to think of heuristics as error-prone shortcuts and Bayesian models as analyses of what people should do. Thus, I am somewhat puzzled by your suggestion that heuristics = representation in a Bayesian model. (If a heuristic is error-prone, then it is not an example of what someone should do.) I suspect that have different things in mind when we refer to "Bayesian" models. Is it possible for you to develop your point a bit more to address my confusion? $\endgroup$
    – Chris
    Commented Jun 28, 2012 at 18:20
  • $\begingroup$ @Chris I don't think bayesian modeling is normative. In general, science typically deals with what is not what should be. Philosophers do try to draw normative claims from the baysian framework. However, the word modeling usually implies that you are trying to reflect what is observed in the real world. If the Bayesian framework only made normative statements and did not have mechanisms to reflect experimental data, then it would have been abandoned by now. $\endgroup$
    – Artem Kaznatcheev
    Commented Jun 28, 2012 at 18:35
2
$\begingroup$

In a paper published recently (actually, today) by myself and my 2 advisers, we analyze results of an auditory perceptual discrimination task, and show that the Bayesian model can be used to explain some aspects of behaviour, but not others. We provide a simple heuristic model that accounts for a wider range of phenomena in that task, such as the imperfect adaptation to non-standard conditions.

This result supports the more general claim that the brain employs heuristics rather than exact Bayesian computations. However, these heuristics are not chosen at random. They are probably developed during evolution, which may explain why, at least in most cases (under common, or 'normal' conditions) perform in a way that may seem optimal, and perform as the Bayesian analysis would suggest an optimal observer/decision-maker should.

Reference

Raviv O, Ahissar M, Loewenstein Y (2012) How Recent History Affects Perception: The Normative Approach and Its Heuristic Approximation. PLoS Comput Biol 8(10): e1002731. doi:10.1371/journal.pcbi.1002731 Full text link

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.