LeabraLayerSpec

From emergent
Jump to: navigation, search
Reference info for type LeabraLayerSpec: Wiki | Emergent Help Browser

Back to Leabra

A LeabraLayerSpec gives much of the information about how each layer is different from the others.

The new default type of inhibition is FF_FB_INHIB -- this will be set when you hit the Defaults button or create a new type. For this type of inibition (see Leabra page for equations), the main parameter is gi -- typically between 1.5 to 2.3 is the best value -- use a higher number for sparser activations and more competition -- typically for larger networks around 2.2 is best, but for smaller networks or lower inhibition lower numbers can work better. This form of inhibition does not use kwta to determine activation levels -- so you will have to tune the gi value separately from the kwta. However, kwta is still important!! -- it is used to determine the scaling of the netinputs to other layers -- see Leabra Netin Scaling

Among the most important things it specificies:

  • kwta How much inhibition is applied, in terms of how many units on average will be active at any time
  • decay Whether activations in the layer decay between trials. Default is to return the units to resting membrane potential between trials.

It does NOT specify how the units within the layer behave, which is specified in LeabraUnitSpec

nor whether that layer is used as an input, output, or internal (hidden) layer; this is specified in LeabraLayer

IMPORTANT: you must accurately set the kwta settings even for input (clamped) layers, because this determines the scaling of the netinputs to other layers -- see Leabra Netin Scaling

For everything else, refer to tooltips for help guessing, or post questions to a list, then update this page with info you found important!