On Tue, Jul 10, 2012 at 3:14 PM, Jarrad Hope <[email protected]> wrote:
> It seems strange to me that NN's aren't effective for dealing with reason, > logic and language, etc - after all our brains seem to be capable of it? > Other discussions on this list notwithstanding, I'd risk to say that nobody has envisioned a non-symbolic AGI. Whether a "grandmother cell" exists or not neurologically, we do seem to use all kinds of statistical tools to manipulate symbols and "logical" categories. A lot of us want to build an artificial scientist, and look what we 've done: we've spent decades looking for the "object" Higgs boson. And I'd dare anyone to assert that she thinks of a wavefunction as opposed of a small ball when thinking of "particles". In fact, they are called and they are particles in the same logical space as gold particles, dust particles etc. We obviously have no idea what kind of computation allows the brain to excel at symbolic tasks (apparently it takes a whole half-brain). We do know that the symbolic universe is ever expanding, also into non-symbolic domains - how would art teaching or choreography do without a domain vocabulary - in my opinion quite badly. Now, there is some magic possible with symbols that possibly goes beyond statistics, they are not called (in certain contexts) production systems for nothing http://en.wikipedia.org/wiki/Production_system So, from a practical point of view it would be very inefficient to get any statistical or NN structure to produce, let's say, english text by outputing sequences of ASCII characters, and we don't have processing power to waste. In fact we may be dabbling into the realm of impossibility with exponential growth of the problem space (what is the chance of character 79 being z if character 1 was A?), unless the statistical system precisely encodes nearly optimal linguistic rules (caveat: our rule-based systems are not very impressive either, as if we have missed something somewhere) >From a theoretical point of view there may be limits for statistical systems to engage in creative pursuits. Already from the get-go we can assign a fixed complexity to a (trained) ANN, while a production system operates more or less on the Turing machine with its infinite tape and time. Of course they are also equivalent in some sense U http://en.wikipedia.org/wiki/Universal_Turing_machine , but a program outputing Pi's digits is so vastly superior to an ANN being trained to do so that we should probably talk of multiple levels of computability just like we speak of http://stackoverflow.com/questions/999121/multiple-levels-of-infinity Finally, Ben's opinion that "The magic ain't in the neurons" is a very optimistic opinion, considering how magical cells, intelligence and the universe as a whole are. AT ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
