Post date: Mar 11, 2017 4:7:0 PM
Final thoughts/conclusions on this classical issue:
Poor performance on train data => low model complexity/underfit => need to increase model complexity by
+ increase model(feature) dim
+ reduce regularization
Divergence between train and test performance => high model complexity/overfit => need to reduce model complexity by
+ decrease model(feature) dim
+ increase regularization
If the performance on both train and test data is still inadequate => not enough data
+ get more data
+ iterate with more epoches/passes on the existing dataset
Amazon puts it best [1]!
[1] http://docs.aws.amazon.com/machine-learning/latest/dg/model-fit-underfitting-vs-overfitting.html