Busy writing thesis (especially MC chapter to give it to you (David) and afterwards Silvia).

Investigation on cluster flags:

• First ideas about training on flags directly failed because for most clusters the flags are 0 -> Network just learns to produce 0 for everything, so this needs downsampling
• Second idea learn the cluster overlap. This works to some extend

• Two problems need mitigation:
  • Also many clusters with overlap value of exactly 0 -> Removed from training data set now to avoid overtraining. 
  • Network output not bound to be between 0 and 1 without non-linear activation in the last layer: Outputs < 0 are mapped to 0 (sometimes incorrectly)
• Ultimate problem: Cluster overlap does not nicely correlate with cluster flags: