DP
Also you can add stuff like differential plasticity to improve your learning
12 days, not 12 months :P
Size: a a a
2019 July 23
BK
Is this part of NN?
Up
BK
12 days, not 12 months :P
Easy stuff, uber has made a package
DP
i dont think it makes sense for just 1800 samples
way simpler to learn every possible connection (fully connected layers. resources allow it) and delete weak connections (he needs a simple model for presentation)
way simpler to learn every possible connection (fully connected layers. resources allow it) and delete weak connections (he needs a simple model for presentation)
LS
Only 80% of 1800 actually to be Train
DP
heavy dropout (50%) can help avoid overfiting even on small dataset like yours
LS
Okay I need to read more on this. Not good with nn
DP
good part about nn is that if there is ANY correlation between the data and answer, it will likely find it.
and it can be simplified by removing all of the connections, untill you have just a hundred non-zero weights left (prediction power will drop. but it will be easier to explain)
and it can be simplified by removing all of the connections, untill you have just a hundred non-zero weights left (prediction power will drop. but it will be easier to explain)
LS
Okay NN will need no feature selection I think ?
DP
ideal input for NN for your task is a float number
LS
Do I need to normalise the data or scale it ?
DP
'feature selection' is a part of data analysis that is considered to be done by NN
DP
Do I need to normalise the data or scale it ?
ideally if you will normalize it so that minimum is 0 and maximum is 1
DP
if you have any data that differs more than 100 times from non-zero min to max, then log of data also helps. blood analysis can give such data
LS
I will go for normalise
LS
Between 0 and 1
DP
if you have any data that differs more than 100 times from non-zero min to max, then log of data also helps. blood analysis can give such data
this might be required in addition to normalization if you have a data like 0.00001 and 1.0 in the same 'column'