[ExI] Strong AI Hypothesis: logically flawed?

Ohad Asor ohadasor at gmail.com
Mon Sep 29 01:02:49 UTC 2014

I think I understand the probabilistic tools you rely on. As you mentioned,
Bayesian approach and assumptions over families of distributions.
The main crunch of the new machine learning theory is to prove generalization
bounds <http://en.wikipedia.org/wiki/Generalization_error> that apply to
*all* distributions.
Say you sampled 100 ravens and 60% of them are black. Can you tell
something about the unsampled ravens, that will hold regardless to the
ravens' color distribution? Or in other words, be true for any unknown
underlying distribution?
The astonishing answer is yes. It is a consequence of a property called
'concentration of measure'. Here
<http://en.wikipedia.org/wiki/Margin_classifier> is an example. Chebyshev
inequality can be seen as something weaker (since it requires the second
moment to be finite and known) that gives you bounds applying to a wide
family of distributions.
I see the present and the future of machine learning focusing on such
approaches, designing learning algorithms to apply to any underlying
distribution. From here, the road to a general purpose AI, it open. All, at
least IMHO.

What is your opinion regarding those universal bounds?

On Sun, Sep 28, 2014 at 12:01 PM, Anders Sandberg <anders at aleph.se> wrote:

> Ohad Asor <ohadasor at gmail.com> , 28/9/2014 4:56 AM:
> Hi all, great to be here :)
> Hi!
> On Sun, Sep 28, 2014 at 12:58 AM, Anders Sandberg <anders at aleph.se> wrote:
> Decades of failure is obviously some evidence
> Why do you think so, sir?
> I was using it in a Bayesian sense: it is information that ought to change
> our probability estimates, but it might of course be weak evidence that
> just multiplies them with 0.999999 or something like that.
> If one thinks that real AI research is only possible now because of
> computational advances or some relevant new insights, then decades of
> failure are very weak evidence. Just like decades of flying failure was not
> really good evidence against heavier-than-air flying since most of those
> approaches lacked the necessary aerodynamic knowledge: it was only after
> that had been discovered the Wright brothers had a chance. However, now the
> uncertainty resides in whether we think we know enough or not.
> One neat way of reasoning about problems with unknown difficulty is to
> assume the amount of effort needed to succeed has a power-law distribution.
> Why? Because it is scale free, so whatever your way of measuring effort you
> get the same distribution (also, there are some entropy maximization
> properties I think). We also have priors which can be approximated as
> log-uniform. From this some useful things can be seen, like that the
> probability of success tends to grow in a strongly convex way as a function
> of resources spent, that neglected domains can be extra profitable to
> investigate even when our priors say they are difficult, and estimates of
> expected benefit given a certain resource spending and our current
> knowledge. See
> http://www.fhi.ox.ac.uk/how-to-treat-problems-of-unknown-difficulty/
> for a start - Owen have a lot of neat results I hope he puts up soon.
> Anders Sandberg, Future of Humanity Institute Philosophy Faculty of Oxford
> University
> _______________________________________________
> extropy-chat mailing list
> extropy-chat at lists.extropy.org
> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20140929/404f53a9/attachment.html>

More information about the extropy-chat mailing list