How to get intuitive understanding which deep learning architecture suits for my problem

I'm working on a research problem where I need to perform classification for coarse prediction in a feature space and then fine grained regression for getting more precise values. I know that this way of regression should work. I also will essentially deal with feature maps.

I am thinking of using a 'stacked hourglass network'. Do I need to identify this by sheer experimentation or can someone intuitively remove some possibilities saying a particular architecture may not be suitable for my problem.

I found stacked hourglass network to upscale and downscale essentially the heatmaps but now am confused with changing the model for sequential classification and regression task. Any clues would be welcomed.

Stacked Hourglass Networks for Human Pose Estimation

Topic convolutional-neural-network regression deep-learning

Category Data Science

1
answered at

"Stacked hourglass" network is an example of end-to-end learning, error gradients are backpropagation throughout the entire network.

Given that you are trying to fit two different types of models, classification and regression, it might be tricky to do end-to-end learning. It might make more sense to build and train one model, then have the trained outputs of that model become the inputs of the next model. Then train the next model.

About

Geeks Mental is a community that publishes articles and tutorials about Web, Android, Data Science, new techniques and Linux security.