I am confused with the analogy of spatial pooling in neocortex. So, basically, as far as I understand, every layer in a column receives some input from sensory organs or from other layers in the column or from other columns.
So, does there exist a set of neurons in every layer (say lower in the layer) which will normalize the inputs (spatial pooling) and send it to other set of neurons (upper set of neurons) which will perform TM algorithm?
Also, In the SPATIAL POOLING video in HTM school (episode 7), something is mentioned that “the output vector of spatial pooling is minicolumns”…I am confused about what is a minicolumn?
2 Likes
The primary confusion is usually between cortical columns and cortical minicolumns. In our older literature (including older episodes of HTM School), we unfortunately refer to minicolumns as ‘columns’.
Maybe not every layer, but definitely one at least. We think probably two.
Not exactly. Spatial pooling creates a physical space in a layer for another computation to happen within distal dendritic connections. In HTM school, I showed how temporal sequence memory can occur in this space within minicolumns, but other computations can also occur. It depends on how the neurons are “hooked up” to each other, meaning where the distal connections are coming from and what they represent. Did you see the Cortical Circuitry episode?
4 Likes