Niels Leadholm, a visiting researcher, discusses some ideas for further research on how to apply the object recognition implemented in Numenta’s 2019 paper Locations in the Neocortex: A Theory of Sensorimotor Object Recognition Using Cortical Grid Cells to images. In particular, the research would explore whether the strengths of object reference frames and grid-cell encoding can be leveraged in an image-based setting.
Thanks for interesting idea! It is not clear enough for me how he create 5x5 grid cell Modul from a MNIST image. We can encode image using grid cells so for his 64 channels he has to use 64 gridcells modules with different scale, orientation and offsets.
Am I right here?
I don’t think he utilized any location information or reference frame kind of concepts here.
The SDRs don’t seem to imply true location information.
This just looks like a convolutional spatial pooler to me.
His approach of a standard CNN + k-WTA looks to me like a general strategy to encode arbitrary stimuli for HTM systems, similar to the “SDR classifier” decoder.
Thanks