Leabra Implementation Reference - Computational Explorations in Cognitive Neuroscience - page 452

LeabraLayerSpec layer-level specifications:

Variable Default Description

kwta How to calculate desired activity level:

.k from How is the actual k determined: USE K = directly by given k, USE PCT = by pct

times number of units in layer (default), USE PAT K = by number of units where
external input ext > .5 (pat q) .

12 Desired number of active units in the layer (default is meaningless — change as

appropriate).

.k

.pct .25 Desired proportion of activity (used to compute a k value based on layer size).

Desired activity level for the unit groups (not applicable if no unit subgroups in
layer, or if not in inhib group). See kwta for values.

gp kwta

What to consider the inhibitory group. ENTIRE LAYER = layer (default),
UNIT GROUPS = unit subgroups within layer each compute kwta separately,
LAY AND GPS = do both layer and subgroup, inhib is max of each value.

inhib group

How to compute inhibition (g i): KWTA INHIB = basic kWTA between k and
k+1 (default), KWTA AV G INHIB = average based, between avg of k and avg of
k+1-n, UNIT INHIB = units with inhib flagsendg i directly.

compute i

i kwta pt .25 Point to place inhibition between k and k+1 for kwta (.25 std), between avg of k

and avg of k+1-n for avg-based (.6 std).

adapt pt Adapt the i kwta pt point based on difference between actual and target pct

activity level (for avg-based only, and rarely used).

.005 Time constant for integrating average average activation.

.a dt

.05 Tolerance around target avg act before changing i kwta pt.

.tol

0 Time constant for integrating i kwta pt changes (0 = no adapt, .1 when used).

.pt dt

.2 Maximum deviation from initial i kwta pt allowed (.2 std).

.mx d

How to clamp external inputs.

clamp

true Whether to hard clamp external inputs to this layer (directly set activation, result-

ing in much faster processing), or provide external inputs as extra net input (soft
clamping, if false).

.hard

.gain .5 Starting soft clamp gain factor (net = gain * ext).

.d gain 0 For soft clamp, delta to increase gain when target units not >:5 (0 = off, .1 std

when used).

Proportion of decay of activity state vars between various levels of processing:

decay

1 Decay between different events.

.event

1 Decay between different phases.

.phase

0 Decay between 2nd set of phases (if applicable).

.phase2

Link inhibition between layers (with specified gain), rarely used. Linked layers
are in layer objects.

layer link

false Whether to link the inhibition.

.link

.5 Strength of the linked inhibition.

.gain