[Insight-users] Jumps or spikes in metric values

[Steven Boyd]

Hello,

When tracking the registration metrics at each iteration, I see there 
are periodic jumps in the metric value.  I'm using a gradient descent 
optimizer, and testing with three different metrics: normalized 
correlation, mattes mutual information and mean squares.

Am I correct to assume that these jumps are due to the changes in the 
step size of the optimizer?  If so, why would a step size change lead to 
a spike in the metric value?  Why are these spikes more prevalent with 
the mean squares metric than the normalized correlation or mattes mutual 
information metrics?

A sample plot of the metrics can be seen here:

http://www.enme.ucalgary.ca/~skboyd/download/data/Metrics.jpg

Regards,

Steve