X-Message-Number: 8639 From: eli+@gs160.sp.cs.cmu.edu Subject: Re: CryoNet #8632 - #8635 Date: Mon, 29 Sep 97 13:08:02 EDT >From: Thomas Donaldson <> >Subject: Re: CryoNet #8631 >Date: Sun, 28 Sep 1997 16:47:44 -0700 (PDT) >The reference is by HT Siegelmann, ED Sontag, SCIENCE (268(1995) 545-548). OK, I looked through this -- not in detail, so I may be missing some fine points. Nor did I look for followup letters... >It turns out that a neural net with real-number weights on >each connection cannot be imitated by a Turing machine. FULL STOP. This analog-map model doesn't really involve real (infinite-precision) numbers, i.e. it's not uncountably infinite. (They formulate it using reals, but they note that the infinite precision is superfluous.) As a result, it can be emulated with a TM -- at the cost, as usual, of some time. (No need to get fancy to make this claim: the brain's parallelism would also take time to emulate on a serial machine.) >The implication that our brain's neural nets might do some things >impossible for a Turing machine bears a lot on the various >controversies that have been going on in Cryonet. Wait a minute. The _Science_ article compares two mathematical models. This has no direct bearing on brains and computers. (Much less on the "various controversies"...) The article calls the model "realizable", but I don't think that means what it seems to mean (i.e. could you build a useful one?). The model requires unbounded precision, which is not available. You may say that the TM's unbounded tape space doesn't exist either, but there's a difference: linear space requires roughly linear resources. If the world can hold the input for the problem, it can be convinced to yield linear-sized working space. Linear precision (which their model requires for a linear-sized job) generally takes exponential resources. If you want to place a point on a line with 20 bits of precision in the presence of 1-mm noise, you need 2^20 millimeters worth of line. If the authors can work around this, their model becomes more interesting. Of course, there are standard "digital"[1] unbounded- precision models too, and guess what? They're faster than TMs... [1] the distinction really is meaningless for countably-infinite models. -- Eli Brandt | eli+@cs.cmu.edu | http://www.cs.cmu.edu/~eli/ Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=8639