Aye Jack, good stuff and a great

mrG — Sun, 11 Jun 2006 10:53:11 -0400

Aye Jack, good stuff and a great start; another question that intrigues me is in multi-layer backprop, when we start to probe the inner-layer weights for abstract concepts, abstracts it is quite possible we humans have overlooked, abstracts that might, dare I say. be useful to a Google AdSense and such? Could be as well that inner-layer weights might betray the "user interests" you are looking for?

propogation through blogs

jackhebert — Sun, 04 Jun 2006 02:58:55 -0400

I know that this is an old entry that I am commenting on - but google finds old entries. I thought that you might find this research I did interesting as it shows that information propogation through blogss can be successfully modeled with single layer neural nets.:

http://cs.washington.edu/homes/jhebert/pageBlog/blogspace.html

The Blogs Know

mrG — Fri, 13 Jun 2003 08:32:45 -0400

Some think of blogging as a one-way street, a soliloquy that struts it's hour on the stage, signifying nothing; given the huge population of blogs of many colours, this is no doubt true in some specific cases, but just as the list of downers and uppers, if you think they necessarily function in a echo chamber of their own making rest assured it's an illusion that only exists for some small number of blogs, and those only for some of the time.

Case in point is when James Bow learned a lesson about Wolves in the process of working on his new novel. James had originally cast the creatures as coyotes, until a knowledgeable reader noticed and point out that these creatures only came to Ontario after the period of his story. So he changed to wolves, but then ran into the question of the effectiveness of the wolf bounties just prior to his period of interest.

Now, before we start thinking that the blog-channel is the only channel for knowledge, it's worth noting how James eventually telephoned preditor biologist Maria DeAlmeida and MNR, but the fact remains, the process started, just like Joey's New Girl there is not only a broadcast channel of blogs into the sea of browsers, but also the returning channel of browser feedbacks into the blog, and the (recent) history of blogging, if you ask me, has been not in the making of more effortless posting (otherwise Textile would be ubiquitous) but in the variety of feedback channels that enable each blog to remain self-correcting, perhaps not on it's own bit of screen real-estate, but certainly in the aggregated space of the sum over all.

The thing is, the feature-poor bloghosts with the author/member constraints, mainstream media with their comment-less format are only fooling themselves: The feedback still happens, you know it does, it just doesn't happen at your house. Ditto for the company intranets and proposed closed-circuit, managed b-blogs -- what they miss from this is self-correction, the bio-feedback that tells them their wolves are the wrong colour; the naiive go away with bad data, the informed go away laughing to themselves, and everybody loses.

In AI circles, we call this technique 'back propagation' and it works like this: You have a row of inputs (say, the media) and you have a row of outputs (say, google and technorati) and in-between these, you have a seething jumble of 'nodes' each with an arbitrary number of inputs and an arbitrary number of outputs, some nodes connecting to the sources, some nodes connecting to the outputs, but all of them connecting to some small subset of other nodes.

Now, an input happens (news), it 'excites' one blog, who excites five others, who exite a few more and so forth, and each node colours the input by it's own personal metric of relative importance on the inputs and the amount of effort it puts on its outputs.

This first naiive model is called the Perceptron and it works fine for things like predicting the weather from a barometer, but it's fixed and static, awkward to train, and it doesn't adapt to environmental changes. This is, I believe, the model at the heart of many commentaries on blogs who believe blogs are individual points reacting in constant reflex according to the powerlaw of their connectivity.

But real blogs are pilotted by humans, and humans do two things: They think, and they talk. Ok, as Bullwinkle said, I never said they can do it well! but look what happens -- they tell their input node, either directly or through an nth-generation grapevine, that the received data is either good or bad. Maybe through comments, maybe only through referrer counts and pagerank, the message goes back that the data was ok as is, or might need some change.

We call this model, Back-Propagation and it's sufficient to keep a single-celled organism alive.

Now, just within AI, we quickly realized our industrial processes are more complex than seeking light and motion, and we realized the juju in successful back-propagation networks was in the weights, the importance each node gives to the forward or backward incoming connections. Sometimes it's good to be an arrogant pompous bastard, sometimes it's good to be a wishy-washy blog-whore, it's up to each node to anneal to locally optimum weights and that, as it turns out, happens pretty darn optimally through Genetic Algorithms ... aka "natural selection".

Hey Little Red!

This is an aside, or maybe not, but there's a phenomenon worth noting about back-propagation networks, and it's usually exemplified with the Little Red Riding Hood problem -- the network is exposed to the cues and the results and learns to discern the wolf, wood-chopper and grandma, but what is most surprising is the way the interior network re-organizes itself in link-clusters around hub nodes (a-listers?) who become abstractions, archetypes if you will, representing we know not what, but being critical to the knowledge and query performance of the whole system.

Could it be Shirkey's powerlaw shows nothing more than how we, the members of the great soup of nobody nodes, by necessity of the backpropagation annealing via genetic algorithms, make a-listers as a means to abstract particular concepts in the blogspace? -- as I've said before, Madonna is only Madonna because we choose her as our zeitgeist, she's our intellectual property. We create our Super-stars as a necessary and natural cognitive scaffolding and no more 'special' than your browser's bookmarks. They are just a place to hang our links, they remember our connections for us and it's only later we go back and identify them as a "thing in itself" the abstract archetype that says whether or not I'm a Pepper and if you're a Pepper too.

Ok, now, wait a minute here. We have many inputs, a few outputs and a vast sea of interconnected knowledge-distilling self-organizing back-propagation nodes, some of those interior nodes organizing into large complex reputation and reporting systems eventually return a result to be presented to some 'intelligence' ...

Seems here, if you ask me, I've just shown how, in a very real sense, the blogspace truly is wearing mankind as your skin.

TeledyN - The Blogs Know - Comments

Aye Jack, good stuff and a great

propogation through blogs

The Blogs Know

Hey Little Red!