[Ben.F.Rayfield]

http://en.wikipedia.org/wiki/Restricted_Boltzmann_machine (RBM) is the best kind of statistical AI data structure and paradigm for learning and predicting based on partial patterns, that I know of so far. Like bayesian networks, it can be used a variety of ways and is a building block of more advanced things.

Here's a video about RBMs "The Next Generation of Neural Networks" by Geoffrey Hinton

The data structure of a RBM is, between each 2 adjacent layers of sizes Lx and Ly, a 2d array of scalars, size x by y, and for each layer a set of bit vars for if each node is on or off. To run the RBM, you set the bit vars of the visible layer and use the edge weights to set bit vars in each next layer, going back and forth between the visible and deepest hidden layer, until it converges on some local minimum that it has learned, like the pictures it imagines in Hinton's video.

Scalability of AI could be extremely expanded by new ways to merge data structs. How might we take 2 such RBMs and merge them without losing what they've learned or taking it out of context? Harmony Search is the obvious simplest answer, which simply takes averages with some randomness as to which side gets more weight, of the edge weights to create a new RBM, and then you can see how it reacts to various patterns of bit vars. But there is an open ended research path of more complex ways RBMs or other kinds of AI data structures can be merged, the next simplest being the possible ways to pair nodes in one RBM with nodes in another RBM. Even if you know each visible node is for a certain pixel and those do not move, the hidden nodes in multiple layers have little meaning relative to eachother between different RBMs, but if permutations of them are considered, a good match may be found anyways or at least converging toward it in evolution of many RBMs merging with eachother in various combinations and ways of merging.

How might we merge RBMs or other kinds of AIs in new ways to keep the most of both of their learned behaviors?

[ranch vermin]

I can never quite get the activation of this one right,  I actually learnt it then I forgot!!!!

That demo Hinton does is awesome, how it plays back continuous animation.   He makes fancy stuff.