I feel this is awesome
For example, I trained a neural network (A) to recognize whether it is a cat, this network does a good job
Then I now want to train a network that recognizes handwriting
But my model is doing poorly because my data volume is small or something
I can remove the last layer of A and add a layer (b) to make the output fit the format
resize my handwriting (X) to fit the input
Then I use the handwritten image to train the W_b of the last full connection layer
We can also get results with good accuracy
Because for a neural network, our first few layers generate points, lines, simple graphs, etc.
As long as it is the same kind, we can take it and use it directly