But it is not clear at all why stories do not approximate Bayesian updating. Stories do allow us to reach far into the void of space which cannot be mapped immediately from sensory data, but stories also mutate and get forgotten based how useful they are which at least resembles Bayesian updating. The question is whether this kind of filtering throws off the approximation so far that it is qualitatively a different computation.
I don’t think we can say that the mutation or loss of stories is very close to Bayesian updating. It may be a form of natural selection, and maybe sometimes the trait being selected for is “truth”, but very often it’s going to be something other than truth. Memes mutate in order to be more viral, and may lose truth on the way.
Stories about big, shocking, horrible events are more memetically contagious and will thus look more probable, if you’re assuming that their memetic availability reflects their likelihood.
Even if stories are selected for plausibility, truth and whatever else leads most directly to maximal reward only once in a while, that would probably still be equivalent to Bayesian updating, just interfered by an enormous amount of noise.
I don’t think you can justify using the word “equivalent” like that. I think maybe you mean “evolution and memetics are similar to Bayesian updating in some ways”. That is not the same thing as “equivalence”. It is not really helpful to take a very specific thing and say that it is “equivalent” to other very very different things, especially if such a comparison does not help you make any predictions.
My culture has a story in it that the Creator of the Universe is going to come down in the form of a man and destroy the world if people do too many things that are said to be bad by a certain book. There is no plausible way in which the process by which this meme has propagated can be explained by Bayesian updating on truth value.
I didn’t mean ‘similar’. I meant that it is equivalent to Bayesian updating with a lot of noise. The great thing about recursive Bayesian state estimation is that it can recover from noise by processing more data. Because of this, noisy Bayes is a strict subset of noise-free Bayes, meaning pure rationality is basically noise-free Bayesian updating. That idea contradicts the linked article claiming that rationality is somehow more than that.
There is no plausible way in which the process by which this meme has propagated can be explained by Bayesian updating on truth value.
An approximate Bayesian algorithm can temporarily get stuck in local minima like that. Remember also that the underlying criterion for updating is not truth, but reward maximization. It just happens to be the case that truth is extremely useful for reward maximization. Evolution did not achieve to structure our species in a way that makes it make it obvious for us how to balance social, aesthetic, …, near-term, long-term rewards to get a really good overall policy in our modern lives (or really in any human life beyond multiplying our genes in groups of people in the wilderness). Because of this people get stuck all the time in conformity, envy, fear, etc., when there are actually ways of suppressing ancient reflexes and emotions to achieve much higher levels of overall and lasting happiness.
In the limit of time and information, natural selection, memetic propagation, and Bayesian inference all converge on the same result. (Probably(?))
In reality, in observable timeframes, given realistic conditions, neither natural selection nor memetic propagation will converge on Bayesian inference; if you try to model evolution or memetic propagation with Bayesian inference, you will usually be badly wrong, and sometimes catastrophically so; if you expect to be able to extract something like a Bayes score by observing the movement of a meme or gene through a population, the numbers you extract will be badly inaccurate most of the time.
Both of the above are true. I think you are saying the first one, while I am focusing on the second one. Do you agree? If so, our disagreement is a boring semantic one.
the numbers you extract will be badly inaccurate most of the time
As its the case with an myopic view on any Bayesian inference process that involves a lot of noise. The question is just whether rationality is about removing the noise, or whether it is about something else; whether “rationality is more than ‘Bayesian updating’”. I do not think we can answer this question very satisfyingly yet.
I tend to think what Cumming says is akin to saying something like: “Optimal evolution is not about adapting according to Bayes rule, because look at just how complicated gene expression is! See, evolution works by stories encoded in G, A, C and T, and most of them get passed on even though they do not immediately help the individual!”
But it is not clear at all why stories do not approximate Bayesian updating. Stories do allow us to reach far into the void of space which cannot be mapped immediately from sensory data, but stories also mutate and get forgotten based how useful they are which at least resembles Bayesian updating. The question is whether this kind of filtering throws off the approximation so far that it is qualitatively a different computation.
I don’t think we can say that the mutation or loss of stories is very close to Bayesian updating. It may be a form of natural selection, and maybe sometimes the trait being selected for is “truth”, but very often it’s going to be something other than truth. Memes mutate in order to be more viral, and may lose truth on the way.
Stories about big, shocking, horrible events are more memetically contagious and will thus look more probable, if you’re assuming that their memetic availability reflects their likelihood.
Even if stories are selected for plausibility, truth and whatever else leads most directly to maximal reward only once in a while, that would probably still be equivalent to Bayesian updating, just interfered by an enormous amount of noise.
Natural selection is Bayesian updating too: http://math.ucr.edu/home/baez/information/information_geometry_8.html
I don’t think you can justify using the word “equivalent” like that. I think maybe you mean “evolution and memetics are similar to Bayesian updating in some ways”. That is not the same thing as “equivalence”. It is not really helpful to take a very specific thing and say that it is “equivalent” to other very very different things, especially if such a comparison does not help you make any predictions.
My culture has a story in it that the Creator of the Universe is going to come down in the form of a man and destroy the world if people do too many things that are said to be bad by a certain book. There is no plausible way in which the process by which this meme has propagated can be explained by Bayesian updating on truth value.
I didn’t mean ‘similar’. I meant that it is equivalent to Bayesian updating with a lot of noise. The great thing about recursive Bayesian state estimation is that it can recover from noise by processing more data. Because of this, noisy Bayes is a strict subset of noise-free Bayes, meaning pure rationality is basically noise-free Bayesian updating. That idea contradicts the linked article claiming that rationality is somehow more than that.
An approximate Bayesian algorithm can temporarily get stuck in local minima like that. Remember also that the underlying criterion for updating is not truth, but reward maximization. It just happens to be the case that truth is extremely useful for reward maximization. Evolution did not achieve to structure our species in a way that makes it make it obvious for us how to balance social, aesthetic, …, near-term, long-term rewards to get a really good overall policy in our modern lives (or really in any human life beyond multiplying our genes in groups of people in the wilderness). Because of this people get stuck all the time in conformity, envy, fear, etc., when there are actually ways of suppressing ancient reflexes and emotions to achieve much higher levels of overall and lasting happiness.
Let’s taboo “identical”.
In the limit of time and information, natural selection, memetic propagation, and Bayesian inference all converge on the same result. (Probably(?))
In reality, in observable timeframes, given realistic conditions, neither natural selection nor memetic propagation will converge on Bayesian inference; if you try to model evolution or memetic propagation with Bayesian inference, you will usually be badly wrong, and sometimes catastrophically so; if you expect to be able to extract something like a Bayes score by observing the movement of a meme or gene through a population, the numbers you extract will be badly inaccurate most of the time.
Both of the above are true. I think you are saying the first one, while I am focusing on the second one. Do you agree? If so, our disagreement is a boring semantic one.
As its the case with an myopic view on any Bayesian inference process that involves a lot of noise. The question is just whether rationality is about removing the noise, or whether it is about something else; whether “rationality is more than ‘Bayesian updating’”. I do not think we can answer this question very satisfyingly yet.
I tend to think what Cumming says is akin to saying something like: “Optimal evolution is not about adapting according to Bayes rule, because look at just how complicated gene expression is! See, evolution works by stories encoded in G, A, C and T, and most of them get passed on even though they do not immediately help the individual!”