On Wed, Jan 22, 2014 at 10:56 PM, Sergio Donal <[email protected]> wrote:
> Recall that you can apply any (linear) algebraic transformation to a > function (I mean, the output of f(x) R^n -> R^m can be thought as a vector > in R^m). > > Regarding kernel methods, it is OK to work with discrete values (e.g., > words, arguments, conclusions, etc.) as long as you define a suitable > distance function (e.g., different arguments could lead to similar > conclusions). After Googling "kernel methods applied to logic" I found: > > http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.115.5487&rep=rep1&type=pdf > http://www.dsi.unifi.it/~passe/papers/aprilchapter.pdf > Thanks, I have read Gartner's book before, but somehow I came to a dead end... The kernel requires us to have a good distance metric, but the distance metric itself requires logic deduction (our original goal is to optimize logic deduction)! For example, the idioms "don't judge a book by its cover" and "clothes make the man" are semantically related, but this relation requires logic inference to deduce. Perhaps kernels can be useful but we cannot have the perfect semantic distance of logic formulas as a given. The job of the kernel is to find logical consequence, and to be able to do its job well, it seems desirable to have the facts organized semantically -- because semantically close premises entail semantically close conclusions. So, perhaps the organization of formulas in space would have to be dynamically adjusted over time... to converge to the perfect values... ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
