This example sets up a simple environment to help you study the behavior of integrating units. There are two main types of unit integration in Lens: input and output integration. In the former case the unit inputs are time averaged and in the latter case the unit outputs are time averaged. Of course, one could use both, but that's a little crazy.

You can learn more about how integration works by reading the first part of the rand10x40.in description.

There is a single input unit in this network connected to two hidden units with a fixed-weight link. There is no bias input to the hidden units. One hidden unit is input integrating and the other is output integrating.

Go ahead and click on the first example in the Unit Viewer. You should see two traces in the graph. The black trace is for the input integrator and the red trace is for the output integrator. The first example has an input of 1 followed by an input of 0. The second has a -1 followed by 1. The other two make transitions between 1 and -1.

You should find that the output integrator makes slower transitions away from 0, but faster transitions towards 0.

Try studying the effect of changing the link weights, gain, and dt. To perform each experiment, click on the first example, store the graph, change the parameter value, then click on the first example again. This way, you can directly compare the curves.

You should find that dt has a greater effect on output integrators but gain and the input strength have a greater effect on the input integrator.

Douglas Rohde Last modified: Tue Nov 21 14:07:40 EST 2000