# Adapting Neural Network to new domain without labels

Is there an approach for the following problem:

Lets say, I trained a neural network on a big dataset for categorizing different fruits in $$k$$ classes. Afterwards I got a nice model, which performs very well.

Now I want to use the model for categorizing fruits in the corresponding $$k$$ classes, as it was planned beforehand. Unfortunately the fruits I want to categorize now are all not ripe yet, but my training set consisted only of ripe fruits. Furthermore I have some pictures of these not ripe fruits, but no labels.

How can I adapt my neural network to these slightly different domain with my pictures of not ripe fruits (and no labels!). Performance on the old task does not matter. The only thing I want, is categorizing not ripe fruits.

My only Idea now is to use virtual adversarial training (VAT) for the unlabeled pictures.

## 1 Answer

• Thank you very much. There seem to be some very promising ideas for my problem. Take my upvote ;) Dec 18 '18 at 16:45
• You're welcome! Yes, these papers looks really exciting... I'll try this techniques soon, we could exchange our impressions and considerations :) Dec 18 '18 at 16:46
• I will definitely report back^^ Dec 18 '18 at 16:46