Leabra Implementation Reference - Computational Explorations in Cognitive Neuroscience - page 446

Variable Default Description

fix savg Sets fixed sending avg activity value for normalizing netin: i.e., ￿ k in equa-

. This is useful when expected activity of sending
region that projection actually receives is different from that of sending layer as a
whole.
.fix false Toggle for actually fixing the sending avg activation to value set in savg.
.savg .25 The fixed sending average activation value — should be between 0 and 1.
.div gp n false Divide by group n, not layer n, where group n is the number of actual connections
in the connection group that this unit receives from (corresponds to a given pro-
jection). Usually, the netinput is averaged by dividing by layer n, so it is the same
even with partial connectivity — use this flag to override where projection n is
more meaningful.
savg cor Correction for sending average activation levels in hebbian learning — renormal-
izes weights to use full dynamic range even with sparse sending activity levels
that would otherwise result in generally very small weight values (equations 4.18,
4.19, 4.20).

tion 2.15: g ek =

.cor 1 Amount of correction to apply (0=none, 1=all, .5=half, etc): q m in equation 4.20:

,where m = : ￿ m (equation 4.19), and ￿w ij = ￿[y j x i (m￿

(equation 4.18).

Source of the sending average act for use in correction. SLAYER AV G ACT
(default) = use actual sending layer average activation. SLAYER TRG PCT =
use sending layer target activation level. FIXED SAVG = use value specified
in fix savg.savg. COMPUTED SAVG = use actual computed average sending
activation for each specific projection — this is very computationally expensive
and almost never used.

.src

.01 Threshold of sending average activation below which Hebbian learning does not

occur — if the sending layer is essentially inactive, it is much faster to simply
ignore it. Note that this also has the effect of preserving weight values for pro-
jections coming from inactive layers, whereas they would otherwise uniformly
decrease.

.thresh

LeabraBiasSpec connection specification for bias weight (bias weights do not have the normal weight bounding and
wt limits settings, are initialized to zero with zero variance, and do not have a Hebbian learning component):
Variable

Default Description

dwt thresh .1 Don't change weights if dwt (weight change) is below this value — this prevents

bias weights from slowly creeping up or down and growing ad-infinitum even
when the network is basically performing correctly — essentially a tolerance fac-
tor for how accurate the actual activation has to be relative to the target.