Hi, The current non-biological CLA classifier has lot of limitations. I have thought about a biological classifier that can make predictions. We wont have to keep task specific tables to look up for future values.
every cell suppose is connected to every other cell that is not in its column. All connections would have weights which get modified based on the patterns falling on the network. After a pattern has fired suppose 2 cells in 2 different columns, next pattern would suppose fire 2 new columns completely(bursting). Then 1 cell each from those 2 bursting columns is selected and connections between those 2 cells and the previous 2 cells that fired for previous pattern would get strengthened. Next time the 1st pattern falls on the network, these strengthened connections will put those two cells in predictive state. Same can be extended to 3rd, 4th etc order of prediction. These weights can be stored in a matrix. So every cell will have its own weight matrix, every value representing the strength of its connection with other cell at that relative place. There wont be multiple look up tables required for multiple order prediction and one matrix per cell is the only fixed memory required by the classifier. If I wasnt totally comprehensible I can make a short video putting it all together. Thanks !! Aseem Hegshetye _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
