How to know or decide on how many hidden layers required for defining predictive model for image classification tasks? Is there any methodologies?
More hidden layers and more neurons in a neural net is helpful in accuracy, but in the other hand it becomes computationally heavy task. So to bring performance we need to reduce layers/neurons and sacrifice accuracy a bit.
As a developer of a model you need to get the perfect balance between performance and accuracy. In other words to tune your model you need to know how much accuracy you are willing to sacrifice for the performance.
Though you will notice after some number of layers if you add further layers you are not getting significant accuracy increase. In that case I will stop increasing the layers.