Why I got better results with ascending numbers of neuron in the hidden layers?

Question

I recently finished a Kaggle contest("MoA Prediction"), I used the method of averaging out the results of different neural networks, and scored 0.01626 in log loss