Can Convolution neutral network train(learn) separately (train different times)?

Question

I am new in Convolution neutral network(CNN). My question is, is there a way to let CNN train separately? For example, at very beginning, CNN only need to learning hardwriting 0 and 1. After the training is good enough(0 and 1 only), now, I want CNN to learn hardwriting 2. So, since 0 and 1 had learned really good, is there a way to let CNN just learning hardwriting 2 only instead of learning everything (0 1 2) from scratch ??

You know, just like human brain, after you already know how to distinguish 0 and 1, we just need to learn 2, then we do can distinguish 0,1,2. instead of clean our memory first,then learning 0, 1,2 at same time. Because that efficiency is so low. Please advise me. Thanks.

Answer 1

This is called curriculum learning , and not only this approach is possible but also it is shown to be quite effective. A few years ago I also came up with similar idea. Basically the idea is, instead of flooding the network with different examples, we can categorize the training data from simple to hard and gradually feed data ordered based on difficulty. After searching for a while, I encountered this paper by Yoshua Bengio, which i think is among the important ones on curriculum learning. Also check out FlowNet 2.0 which is a network trained with this approach.