Raw Pixels | Center - Surround | Oriented Gradient | |
---|---|---|---|
SEV:
Edge Task |
traces well;
learns very fast |
traces well;
learns fast |
traces well |
SEV:
Line Task |
X | X | X |
FD:
Edge Task |
traces well | learns fast | traces well;
learns fast |
FD:
Line Task |
traces ok | traces well | traces well; learns fastest |
In one specific case, the SEV output with raw pixel inputs on the edge-following task was the fastest learning system, reaching its learning objective in 200 epochs. With the other two input representations, several thousand training epochs were required. The drawback of the SEV output model is that it works only on edges, however, and fails miserably when learning to follow a line.
With the FD output model, both tasks could be learned well. For the edge-following task, the center-surround and oriented-gradient filters both reached their learning criteria in 100 epochs vs. 800 epochs for raw pixel inputs. Equivalent improvements were evident for the line-following task.
It is interesting to note that the SEV outputs with raw pixel inputs was the fastest learning, but most specific combination tested. In a way, they are custom fit for each other. On an edge, the pixels on one side will all have low values and high values on the other; it should be easy to learn a smooth mapping onto [0,1] in this case.
Using the FD output model, the network could learn both tasks, and in
this case the preprocessed inputs yielded faster learning. This neurologically-inspired
combination was a more general learning mechanism, even though somewhat
slower in learning overall.