What can be done to correct for sampling bias introduced from (noisy) training data while training a DNN?

Question

The obvious solution is to ensure that the training data is balanced - but in my particular case that is impossible. What corrections can one perform in such a scenario?

I know that my training data is heavily biased towards a particular class, say, and I cannot change that. Moreover, the labels are very noisy. Conditioned on this piece of information, is there anything I can do by tweaking the training process itself/ something else, to correct for the bias in the training data?

The data comes from an experiment (from an electron microscope), and I cannot collect more data. It's always going to be biased in this way, so alternatively-biased is also not an option. I'm sorry that I'm unable to provide any more details due to confidentiality.

Answer 1

I feel like from the information your giving (some sort of biased data) you can't get an answer as robust as you'd like (what algorithmic changes can be made).

In general, the reason these methods like DNN's work is that they learn from the data. What you train it to do is what it is capable of, and there's little one can do to 'balance' it to classes of data it just never sees. It's like training someone to do algebra then giving them a trigonometry test. It's all math, sure, but you just never can expect much without the proper learning.

That being said, you should perhaps look at other methods to work with this data, or to approach the problem. Given that you cannot collect unbiased data and that you can't explain more due to confidentiality, I really doubt anyone here can help you that much.

I can at the most point you to this article : "Classification on Data with Biased Class Distribution".

And suggest that perhaps your current approach may not be the most appropriate given the unfortunate circumstances.