Information Technology Reference
In-Depth Information
Variable
Description
thr
Threshold for firing as used in computing activation from v m , as a state variable
so that it can be adapted as a function of activation over time (not done by default,
done if adapt thr.t dt > 0 in UnitSpec — if adapting, treated like a weight
and only initialized with weights).
LeabraUnitSpec unit-level specifications:
Variable
Default
Description
0, 1
Range of activation for units: Leabra units are bounded between 0 ( min )and1( max ).
act range
bias con type
Type of bias connection to make: almost always LeabraCon.
The LeabraBiasSpec that controls the bias connection on the unit.
bias spec
The activation function to use: NOISY XX1 (default), XX1 (not convolved with
noise), LINEAR ( act is linear function of v m above threshold (0 below thresh-
old)), SPIKE (discrete spiking).
act fun
act
Specifications for the activation function:
.25
The threshold value ￿ in equation 2.19: y j =
. Note that the units
have their own adjustable thr parameter used if adapt thr.t dt > 0 .
.thr
￿ [ Vm ￿ ￿] +
￿ ￿]
.gain
600
Gain of the activation function ( ￿ in equation 2.19: y j =
).
￿ ￿]
.005
Variance of the Gaussian noise kernel for convolving with XX1 function in
NOISY XX1.
.nvar
￿ ￿]
￿ [ V
Specifications for the discrete spiking activation function (SPIKE):
spike
3
Spike duration in cycles — models extended duration of effect on postsynaptic
neuron via opened channels, etc.
.dur
.v m r
0
Post-spiking membrane potential to reset to, produces a refractory effect and con-
trols overall rate of firing (0 std).
10
Gain for computing act eq relative to actual time-average spiking rate ( ￿eq in
equation 2.18: y eq
j
.eq gain
).
.ext gain
.4
Gain for clamped external inputs, multiplies the ext value before clamping,
needed because constant external inputs otherwise have too much influence com-
pared to spiking ones.
adapt thr
Adapting threshold specifications (not used by default):
.a dt
.005
Time constant for integrating activation average act avg : y avg
j
.
.01
Minimum avg act, above which no adaptation occurs.
.min
.35
Maximum avg act, below which no adaptation occurs.
.max
0
Time constant for integrating threshold changes when avg act is outside of
min,max bounds.
.t dt
When set to 0, adaptive threshold mechanism is turned off.
To use, set to .1.
.mx d
.04
Maximum amount to change threshold — even if avg act is not restored to
min,max bounds by changing threshold, don't keep changing beyond this devi-
ation from orig threshold.
Search WWH ::




Custom Search