mirror of
https://github.com/BlackLight/neuralpp.git
synced 2024-12-26 19:35:12 +01:00
12 lines
812 B
Text
12 lines
812 B
Text
Sometimes the training phase of the network breaks in the middle. It happens
|
|
because the synaptical weights are initialized with random values, and
|
|
sometimes updating them causes those values to become >= 1. This makes the
|
|
output values of the network diverge instead of converging to the desired,
|
|
expected values. The library recognizes this behaviour, and when a weight
|
|
become >= 1 throws an InvalidSynapticalWeightException. So far there's no
|
|
way to prevent this odd, random behaviour. The network implements the usage
|
|
of an inertial momentum coefficient to avoid strong oscillations in the
|
|
training phase, in order to make this phenomenon rarer, but also using this
|
|
mechanism there's a possibility ~ 10% of getting a diverging network, and
|
|
so a training phase broken by an InvalidSynapticalWeightException.
|
|
|