Playpen: Learning Demo
Purpose
This demo illustrates the Playpen learning algorithm. Learning
has two phases, a positive phase in which both input and output
units are clamped to a training pattern, and a negative phase in
which only input units are clamped. In an auto-associative
task, as in this demo, there is no distinction between input and
output units, and during the negative phase, units are randomly
selected to be clamped. Because the units are object units,
both activation and phase angle may be clamped. During the
negative phase, these are clamped independently.
In the demo, you will see how the initial random weights in
the network change as the network is trained on 4 simple
patterns. Following training on 10 or 20 repetitions of the
patterns, the network easily completes partial test patterns.
The Network
The network consists of a single layer of 16 units arranged in a
square. All units are connected to one another, and the initial
weights are random. There are four training patterns, each
consisting of an activated horizontal line within the square;
the phase angles of all of the activated units are the
same.
What You Will See
- Note the initial random weights connecting the units, and
compare these to the weights following training. The
trained weights should be positive within the horizontal
lines and negative elsewhere.
- When you click on "Train", there will be 10 repetitions of
all of the patterns. For each pattern, you will see a
positive training phase in which units in a horizontal line have
their activations and phase angles clamped, and other units
are clamped off (though with varying phase angles). This is
followed by a negative training phase in which only a
portion of the same training pattern is clamped. Some units
may be completely unclamped; others may have only their
activations clamped; others may have only their phase angles
clamped; other may be completely clamped.
- When you do "Test pattern", the network is presented with
a pattern in which a single unit is clamped. If the network
has been trained sufficiently, it should complete the
horizontal line in which the unit occurs (with the same
phase angle), and the other units should remain inactive.
Performance on these patterns should be much improved after
10 repetitions of the patterns, and it should continue to
improve if you train it longer (by clicking on "Train" again).
Return to Playpen Technical Report 1
Michael
Gasser
Last modified: Tue Jun 24 01:28:05 EST