1805a NEW Problem
- Jan 10, 2014
The sudden drop or change should not be seen as a problem or even necessarily significant. The sudden drop may well indicate that the NN has changed direction or composition or has followed a new learning path, there are numerous other possibilities within a network, that is why it is complex, but the important thing to focus on is the result, that your changed network gives you, in its predictive capabilities, not its aparent diection. Sometimes these indicators can be a real 'false positive' so don't get hung up on them or you could find yourself wasting an awful lot of effort and time. NN are wonderful for their aparent obsequiousness just don't get fooled. It's what they are able to do that matters.
As to 'Import Importance' ".......Does not show anything......." do you mean the page is blank?
From: To: Date: Jan 10 2014, 10:44 AM Subject:a NEW Problem
I have been using the Neural quite successfully UNTIL a new Problem has arrived as follows:
Im running a network (which has been running successfully) but now a NEW dataset I am importing (same type of data I have been running succesfully). However the AvgTrainingError starts at .1225 and stays there until 1500 of 2000 Learning Cycles are then and then it DROPS suddenly to .0325 ….
Also InputImportance does NOT show ANYTHING !! (which is important for me to see)
So, there has to be something in this Input Dataset that is making this erratic behaviour BUT no Clues to what it might be….
This email is free from viruses and malware because avast! Antivirus protection is active.
- << Previous post in topic