[FoRK] Q re: ConceptNet (also FluidDB)
J. Andrew Rogers
andrew at ceruleansystems.com
Wed Oct 21 20:38:39 PDT 2009
On Oct 21, 2009, at 12:24 PM, Stephen Williams wrote:
> I view a lot of the semantic technology as A) a better way to
> organize information for straight business purposes and B) often the
> best form of bottom-up feeder information that could feed into
> knowledge and probabilistic graph reasoning systems.
A (the?) big problem is that even if A) is true in theory it is not
true in practice. The organization exposed to the user has little
relation to the organization under the hood. That impedance mismatch
is usually very expensive. Explicitly structured data typically
matches common use cases pretty well, a cheap "80/20 rule" optimization.
> I'm glad you now see a mapping / equivalence. I would really like
> to focus on proving what my intuition tells me, but have several
> more pressing concerns first. Someone will eventually map the
> equivalence of NN with probabilistic relational graph reasoning
> (Bayesian / Markov with imprecise reasoning, etc.) and both will be
> A) unified or B) both be stronger as a result.
A lot of work gets done that makes the assumption that classic NN
models are inferior approximations of high-order Bayesian models. It
may just be a conjecture that is commonly assumed, though it does
appear to be obviously true upon non-rigorous inspection. Anyone can
add weights to the edges and vertices of a graph, the real trick is
deciding where to not put edges even when you have a nominal reason to.
I've long wondered if there is a relationship between Braess' Paradox
and pruning algorithms for graph-like computational models but am too
busy/lazy to spend much time thinking about it.
More information about the FoRK