How to proceed when choosing a neural network architecture with a small data set?

Question

Neural networks

Good afternoon. There is a small amount of data (no more than 10-20 pictures for each class, about 40 classes). What architecture for the network should be tried, for the optimal size / quality. I thought to take a pre-trained xception, but, as I understand it, it will be too large for such a set + only color pictures for input, I use bw

Answer 1

Perhaps it is worth deciding through metric learning, and not head-on classification. Well, or still get more data, it's clearly easier.