Very similar to the situation in this problem: What is neural network good accuracy but I don't have enough reputation to comment so I'm asking here again.
The NN I implemented has 9 inputs, 15 hidden units in 1 layer and 1 sigmoid output. The dataset I was using is about (1) if the annual income of ~50K people $\geq$ or $<$ a certain amount and (2) their education, gender, ethnics, etc (total raw features < 20).
I tried altering the regularization coefficient, no. of iterations, adding 1 more hidden layer (a lot slower smh) and even tried swapping the training and test sets but the 0-1 accuracy always topped at around 83%. Is the networks underfitting my data? What else can I do to improve my NN?
Straightforward logistic regression gives around 81% accuracy, so I bet NN is the way to tackle this prediction problem.
Btw, yes, I'm learning this topic from Prof. Andrew Ng as many others did.