I emerged.. but scathed….. from an ignorance session where Ned Wakeman and I tried to discuss web3-4 “fractal semantic data integration, auto-ontology generation” with too few facts to share between us. it came to this:
To reason over what action to take next, for example which “merchandising offer” shall I place before this customer NOW, do I need a model of the world’s knowledge or do I just need to spot patterns ad-hoc. Today, it was put to me that models are for the birds and that they will never be fresh, complete general enough to be useful. This is not a new view! MIT robotics labs have abandoned some time ago, the attempts to give their robot bugs models of the world, and have instead allowed them to devise, from simple environmentally learned rules, how they should function. Google itself does remarkably well at answering questions without a “model of knowledge” and speech recognition improved remarkably when all the NLP/linguists were fired! If we look at the brain, there seems to be no place where a world model is stored and as we explore our daily lives, we don’t seem to search or querie a central “world model data-set” in our heads. So no models required it seems – just agile stats.
However, True knowledge has made remarkable progress in modeling knowledge and making it useful. We should expect to see that by “blending data” ceverly, we can gain a power law efficacy where 2 facts plus another two facts can be added up to 6 facts. Ubisense is a stunning company selling ultra-fine-grain location tracking systems, with in-building models and fixed UWBand infrastructure, they can build unparalleled systems that would just not be practical with ad-hoc systems. Furthermore, we all know that in signal processing, it’s far far easier to pull a signal from the noise if you have an idea of what you are looking for WCDMA etc etc. Synature has used this to optimise advertising copy in a way that Google could never do by just crunching data raw.
So back to it – as the number of “semantic deals” rises – do we need knowledge structures for the semantic web, ontologies and models? or does the phrase auto-ontology mean something and/or will baysian statistics render our models irrelevant. Where would you put your efforts and me my money?