Does NuPIC allow negative weights for spatial pooler and/or temporal memory?
If so, is that biologically plausible?
HTM does not deal in weights like other artificial neural networks. Synapses are modeled as Boolean connections which have an associated permanence (positive float between 0 and 1). Permanence values above a configurable threshold are connected. Permanence values below the threshold are not connected.
1 Like
As i understood from the code, that’s the way for temporal memory, but is that the same for spatial pooler?
Yes, SP also models synapses in the same way. If you haven’t read it yet, the BAMI chapter on SP is a good way to get a high-level view of how the SP algorithm works. The first HTM School SP video around 9:20 does a good job of explaining this as well.
2 Likes
I’ll be working more on the currently incomplete SP section of Building HTM Systems when I get back from OOTO.