This is a form of Spatial Pooling except with a third dimension added. Pooling layer are stacked in to "columns." Cells can connect to other cells (pick a column, then a layer of the column.) The output SDR is made from checking each column to see if enough cells in the column are active to make the column itself active.

Definitions

Receptive field: the set of inputs a mini-column is receptive to; it will only consider input from items inside the receptive field.

Distal input: input which comes from sibling neuron clusters; supposed to detect and provide context for inputs.

Proximal input: input which comes from sensors.

Bursting: When a column has no active cells and all cells are forced active.

Short version

• BAMI chapter on spatial pooling

Temporal memory works like spatial pooling with some differences:

There are layers to temporal memory; instead of a column being a single bit, a column contains cells underneath which are also their own bits.

A cell is active if enough connections are active.

Cells are connected to a subset of other columns; however they are connected to a random layer of that column.

Each cell in a column's synapses are tested; if enough are active then the whole column fires ("predicts" the output.)

• --

learn sequence of active columns from spatial poolers

predicts what outputs will fire next time cycle

find which cells in active columns are active on this time step

choose a set of these cells to put in a predictive state

cells in active columns sum their connection signals and if they are above a threshold then the cell becomes predictive