Can I add data, that my neural network classified, to the training set, in order to improve it?

Question

Let's assume the following: I successfully trained a neural network on a classification task, it performs well, also on unseen data.

Now my idea is: If the neural network obtains new, unseen data and classifies it, can I add this data to the training set? Via increasing my training set, I want to improve the neural network's performance, as it will always learn new examples, and not stay the way I trained it.

The main problem that I see with this approach is the risk of misclassified samples. They will be rare, but they might occur. Will a few misclassified samples in the training data badly harm my network?

I know, this question is pretty case-dependent, and I don't expect a finite answer. But I didn't find anything to it online, so I'm glad about every hint, if that idea might work out or not.

Answer 1

Your strategy could lead to overfitting, and I would discourage it. It might reduce your training error (or validation error if you put similar data in the validation set), but the error reduction would be misleading -- all your neural net would need to do to look like it improved is repeat what it has already done, which should be easy. At best you will only reinforce what the net already knows.

Answer 2

No, this is not a good idea. Consider a NN with no hidden layers and one output neuron. This is exactly logistic regression.

Consider running logistic regression with two points: -1, and 1, where the label for -1 is 0 and the label for 1 is 1. If logistic regression is performed with some weight penalty, then it will predict positive for all inputs greater than 0 and negative for all values less than 0 -- but not with perfect certainty.

Suppose the test set contains points -0.1 and 0.1. The LR model will classify these as negative and positive respectively, and then if you add these to the training set, the network will get much more confident about classifying samples near 0, even through it shouldn't.

In fact, there could be a data point 0.2 which is negative. This point would be mistakenly classified as positive by the network, but with the augmented with the additional classified test data, the network gets much more confident about the wrong decision.