As we described in the sections above, time is the cue that the visual system uses to achieve its invariance properties. Therefore, the training input to the system is simulated movies of the training images. I used only straight line motions of the images for training this version of the system.
As explained above, the level 1 regions learn most likely sequences of the input movies. Examples of these most likely sequences can be found here. ***
Coincident sequences of level 1 becomes the intermediate object concepts at level2. Click here for examples of such Level 2 concepts.
Once we have the intermediate level concepts, the conditional probabilities are learned by counting the number of occurences of the level 1 inputs for each particular level 2 conecpt. This is expressed as a conditional probability matrix P(x|u) where u's- the level 2 concepts- correspond to the rows and and x's the level 1 patterns correspond to the columns. At the end of the training session, each row is normalized so that the elements of each row sum to 1.
For this version of the system simulation, I handpicked a set of most likely L1 sequences as representative and used them in for the simulations. There are 13 such sequences.
Although I discussed obtaining the condtional probability of each individual L1 pattern, actually what needs to be done is to learn the transitions of one pattern to another under a context. This is still under development both theoretically and in simulations. Please read the tech report for more details on this.