Transfer learning (the method when the training data is poor)

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

Guess you like

Origin http://43.154.161.224:23101/article/api/json?id=324927847&siteId=291194637