Skip to content

Latest commit

 

History

History
48 lines (48 loc) · 2.14 KB

notes.org

File metadata and controls

48 lines (48 loc) · 2.14 KB

Temporal memory

neuron-layer

predictive-state

The predictive state of the neuron is handled as follows: if a dendritic segment receives enough input, it becomes active and subsequently depolarizes the cell body without causing an immediate spike. Neurons in the predictive state (i.e., depolarized) will have competitive advantage over other neurons receiving the same feedforward inputs. Specifically, a depolarized cell fires faster than other nondepolarized cells if it subsequently receives sufficient feedforward input. By firing faster, it prevents neighboring cells in the same column from activating with intracolumn inhibition.

neurons with input above threshold becomes predictive

neurons in predictive state (in previous step) are prefered in the selection during activation

learning

The lateral connections in the sequence memory model are learned using a Hebbian-like rule. Specifically, if a cell is depolarized and subsequently becomes active, we reinforce the dendritic segment that caused the depolarization. If no cell in an active column is predicted, we select the cell with the most activated segment and reinforce that segment. Reinforcement of a dendritic segment involves decreasing permanence values of inactive synapses by a small value p and increasing the permanence for active synapses by a larger value p+.

Don’t use independent SDRs for each layer, reuse instead.

rename to-sdr to something more reasonable

SpatialPooler

Connections

initial pernamence

max synapses per segment

connected pernamence

learning radius

neuron

index

column

hash?

synapses?

synaps

links cell and segment

permanence

segment

creates synapses

stores last used iteration

column

holds cells

single proximal dendrite

get least used cell

ComputeCycle

returns predictive cells

ProximalDendrites

many synapses

DistalDendrites

just one synaps?

Pool

Connections

maps segments to cell

Use 16 bit numbers everywhere

Reinforce most active segment if active neuron was not predictive.