Playpen Demos

Each Playpen demo begins with a set of buttons in a panel. Clicking on these buttons runs the demo. Graphical representations of the units in the network appear in windows which are created when the network is created.

Graphical Conventions

Layers
Each layer of units in the network is represented by a window. Squares within these windows represent the units in the network, but most windows will appear completely black at first because the units begin at their resting activation, which is 0 for most units. When something is happening in the windows, they pop in front of the main browser window, but you may want to shrink your browser and move the layer windows to a place where they do not overlap with it.
Units
Within the layer windows, there is a square for each unit in the layer. For layers of relation units, the square are arranged in groups of five, surrounded by a magenta border, each representing a single relation unit. Within the group, the square in the center represents the bias unit, the two upper squares represent the RU interface, and the layer square represent the OU interface.
Activations
Within the layer windows, there is a square for each unit in the layer. The color of the square signifies the sign of the unit's activation; green for positive, red for negative. The brightness of the square depicts the magnitude of the activation: maximum brightness corresponds to the maximum activation of the unit, and black corresponds to an activation of 0. When a unit's activation and phase angle are clamped, the corresponding square is surrounded by a white border. When only the activation is clamped, the border is gray.
Phase angles
In the square representing each object unit, there is a blue triangle; the direction in which this triangle points signifies the phase angle of the unit. The brightness of the triangle represents unit activation, just as the brightness of the background color in the square does. When a unit's phase angle, but not its activation, is clamped, the corresponding square is surrounded by a red border.
Weights
To observe the weights into a unit, put the cursor over the square representing that unit, and press and hold down any mouse button. Weights are depicted as activations are, green for positive values and red for negative values, and brightness signifies magnitude. One difference is that weights do not have maximum values, so the weight magnitudes are first "squashed" into values between 0 and 1 before being converted to brightnesses. A gray square indicates that there is no connection from that unit to the one where the cursor is. For relation units there are several additional conventions. For weights out of a relation unit into an object unit or another relation unit, you will see the relation unit square divided in half; the left half represents the weight from the left arm of the relevant interface, the right half represents the weight from the right arm. When the cursor is over a relation unit, the weights from object units (to the OU interface) contain a small square if they are connected to the left arm of the relation unit. When units have delay connections into them, there is more than one weight into them from a given source unit. To see the different sets of weights in such cases, don't release the mouse button. You will see the weights with delay 0 first, then the delay weights; as long as you hold down the button, the program will cycle through the weights.

Buttons

Each demo has up to four buttons. Their functions are as follows:
Create net
Clicking on this button creates the network. This may take some time; it will be faster if you have already loaded one of the other demos because they demos share many Java classes. When the loading rocess is finished, you will see one window for each of the layers in the network. You must click on this button before any of the others.
Random activate
Clicking on this button first causes random activations and phase angles to be assigned to all of the units in the network. After a pause, the network then settles; that is, randomly selected units update their activations and phase angles until the changes do not go beyond a threshold. Depending on the network and your computer, this may take quite a long time, but you can interrupt the settling if you have seen enough by clicking on the button again, which will now be labeled "Stop".
Test pattern
The network may be provided with one or more test patterns, inputs which activate or clamp particular units. Clicking on this button causes one of these patterns to be selected and input to the network. Then, following a pause, the network is allowed to settled, as with the "Random activate" button. One difference here, however, is that clamped units do not change their activation (or phase angle, depending on what is clamped), when they update. As with "Random activate", you can interrupt the settling by clicking on the button again. When there are delay connections in the network, it is allowed to settle repeatedly for a fixed number of updates.
Train
The network may be provided with one or more training patterns. Clicking on this button causes the network to be trained on eight repetitions of the set of training patterns. You will be able to observe the positive and negative training phases for each pattern. During each phase, certain units are clamped, and the other units in the network are allowed to settle. You will not see the weight updates which take place during learning directly, but you can examine the weights into units before and after training. As with "Random activate" and "Test pattern", you can interrupt the process by clicking on the button again.

Specific Demos


Michael Gasser
Last modified: Mon Jun 23 23:50:08 EST