Pellicano, Bayes, autism
Elizabeth Pellicano and David Burr When the world becomes ‘too real’: a Bayesian explanation of autistic perception Trends Cogn Sci. 2012 Oct;16(10):504-10. doi: 10.1016/j.tics.2012.08.009. Epub 2012 Sep 7.
We suggest … that attenuated Bayesian priors – ‘hypo-priors’ – may be responsible for the unique perceptual experience of autistic people, leading to a tendency to perceive the world more accurately rather than modulated by prior experience.
What is particularly unsettling for autistic individuals is … the unexpected and unpredictable nature of external events. We suggest that understanding how perceptual systems deal with uncertainty is key to explaining atypicalities in autistic sensation and perception.
[H]ypo-priors in autistic perception may lead to difficulties in using information from the remote past to drive expectations about incoming sensory signals.
As a mathematician I naturally lean to evolutionary Cybernetic views of human perception, hence it seems quite reasonable that:
- Different reasoning habits are suited to different situations.
- Populations will be more sustainable if they have a range of habits suited to the different situations that they may face.
- The treatment of priors should vary between individuals, so that they range from being hyper-adapted to the current situation, extracting the maximum value, to being less efficient but more sustainable. That is, they trade-off short-run and long-run effectiveness differently.
Jack Good discussed discounting the prior, so that where the paper cites Bayes’ rule, P(S|I) is replaced by P(S|I)1-d, where d is ‘the discount factor’. Thus d=0 is what the paper supposes to be normative, and d > 0 gives what the paper calls a hyper-prior, with d=1 being where the prior is not used at all. Good’s point is that d=0 is only appropriate when you are sure that the situation is completely regular. Using d =1 is appropriate if you know that the situation is very novel. Most of the time a small positive d is appropriate, even for most settled adults.
Suppose, for example, that there has not been a run on the banks for 10,000 days. A normal (d≈0) person might suppose that P(run today) ≈ 1/10,000, and hence not be too worried if they see a queue outside the local bank. An autistic person (d≈1) would see the queue and be worried, thus tending to precipitate a run even if the original situation had been innocent. One might argue that the autistic person was ‘wrong’, and that the normal person would have been ‘right’, if not for the autistic reaction. But Good argues that for many real-world situations, d=½ is an appropriate default, and one should decrease d for situations that are unusually regular, and increase it for situations where one ‘expects the unexpected’.
My deduction from the paper is that normal people behave as if there is a lot of regularity, whereas autistic people behave as if there is less. Thus I would expect normal people to do better in situations where there is actually a lot of regularity, autistic people to do better where there is radical instability. Thus in what I would regard as normal countries (like the UK prior to 2007) I would expect normal people to be doing well in finance, with autistic people being marginalised. But in a crisis autism (or at least, heavy discounting of priors) would seem to be an advantage. Better is to be able to adjust one’s discounting to suit the situation. While “difficulties in using information from the remote past to drive expectations about incoming sensory signals” is obviously a ‘bad thing’, so too is “using information from the remote past to drive expectations about incoming sensory signals” when the situation has actually moved on.
Jack Good also notes the dependence of probability judgments on context, C, so instead of a universal P(S|I) he considers P(S|I:C) for different C. One way that hyper-priors might arise is if, instead of discounting, one uses a whole-population context instead of a more appropriate context, or otherwise uses a less rich context. Again, the autistic approach is more appropriate if the more specific context cannot be relied upon, as when you are watching a magician. While autism may be a ‘bad thing’, so too is the inability to see what is really there even when you know it is a trick, as in the paper’s optical illusions.
More generally, if you are aware that you are in a particular situation where ‘you cannot trust your own eyes’, then you should discount what you think you see, or consider alternative explanations. Many supposedly normal people seem poor at this, even after illusions are explained. But what I take from the paper is that autistic people are no better at this: their differences seem to be ‘hard-coded’: not under their control.
Finally, from the paper the merit of under-discounting or being too specific about the context is that it tends to give more stable perceptions as more data is obtained, and often simplifies the alternatives to be considered. This does seem desirable, but it is not the only way to achieve this. Particularly in an organisational setting, one effective approach to reasoning is to identify ‘the most probable’ (d=0, most specific context) explanation, but then to consider broader possibilties (discounting priors, or increasing the scope of the context) to identify significant alternatives. If there are too many alternatives for a reasonably broad range of possibilties, then one may need to take action to simplify the situation. On the other hand, if one tends to discount the current nsituation more than others, it may pay to continually create uncertainty, to give oneself an advantage. Instead of having simple probability assignements, such as P(danger) = 0.01, one ends up with a richer appreciation, such as:
P(danger:C) = 0.01 based on our prior experience, C, but
P(danger:C’) = 1 if they are out to get you.
To me, then, the problem is not so much the merits of broader versus narrower priors, but the merits relative to the situation and other aspects of behaviour.
My notes on: