Home>

I don't know how complex the layer structure should be for deep learning.
How should we consider the layer structure?
Is it studying theory and trial and error?

The contents of deep learning are: "Learning the image of a circle like an attached image and the diameter of each image through deep learning so that the diameter of the circle can be measured automatically." is.

I understand this to the extent that it would be better to use CNN for learning images in the deep learning of regression.

In addition, 6000 pieces of data are generated, 10% is used as test data, and it is tried with a slightly modified layer structure (P.87) that appears in the book called Intuitive Deep Learning,
loss: 5.4620-acc: 0.1720-val_loss: 5.8723-val_acc: 0.1713
It is a result.

  • Answer # 1

    I will study the literature carefully about how to determine the layer structure.

  • Answer # 2

    For a task of this level, it seems that it is sufficient if the parameters are extracted by Hough transform without using CNN.

Related articles