I have a basic question that that the diagrams above tend to highlight. The diagrams show the columns of cells extending across all layers. My understanding of the Temporal Memory is that it is modeling 1 layer. When an input is not predicted and all the cells in a column are activated, is that just for 1 layer, or is the entire column across all the layers activated which would mean that the TM is really a model of all the layers?
Yes, the TM is only modeling one layer within a region (2/3). There is a difference between “mini-columns”, which exist within layers, and “columns”, which are larger and extend across all layers.
Current HTM implementations incorporate only mini-columns. The idea of “columns” is a research topic. If you would like to ask questions about this, I suggest you do it in #htm-theory.
2 Likes
Minicolumns actually do span layers, although we currently model just one layer. I think of cortical columns as defining which cells share FF receptive field and cortical columns being the collection of minicolumns within the inhibition radius. This might not be quite right but it is how I think about it currently.
5 Likes