> Multiple cells per column offer variable order memory, while one cell per > column offers only first-order memory. However, first-order memory is useful > for static invariant pattern recognition, since it doesn't maintain long > context. This might be ideally suited for the image categorization problem. > > From the whitepaper: "We have further observed that a region like this > exhibits stability to translation, changes in scale, etc. while maintaining > the ability to distinguish between different images. This behavior is what is > needed for spatial invariance (recognizing the same pattern in different > locations of an image)." > > So maybe try a one-cell-per-column region!
This is exactly what got me excited. But I can't understand how that works and why would that work. Example. I feed an image to a first order HTM. I get columns activated. So HTM learns which columns represent this image. Now each cell (1 per column) is connected to adjacent other cells as well. So when a rotated image comes in, it activates new columns. Cells in those new columns have synapses to cells in columns from the first image. I.e. their prediction is the original variant of an image. But unless it is implicitly specified, those synapses have no way of figuring this out. Again this is a problem of showing an HTM every possible size and angle to ensure invariance. I wish I'd understand where the invariance comes from, without exposure to every possible combination which can be a huge number of combinations. _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
