Aug 24, 2017 · 1 min read
I have a problem. I use CNN network with 2 layers for NLP task. But the model is underfitting severely. I tried small data sets with 5–10 epochs, but loss did not reduce (by the way, I use triplet loss). Should I add more hidden layers or what? I guess that the problem is my initialization function — truncated_normal from TensorFlow, I have no experience in choosing best init function. What function should I try to overcome undefitting?
I use same architecture that I found in literature, so probably going deeper with more layers would not help