Neural network model that ouputs how similar a new input is to the previously trained ones for curiosty-based learning techniques

Question

I don't find where to ask this or if this is too basic or complex to be solved, here is my question:

Is there a NN model like a CNN where you would recursively input images and get a single output telling how similar is that new image compared to the ones processed in past batches?

An example would be: You feed de model with images of furry animals till some point the last inputed images give a value of 1.0 as they are similar and share features, then you input a picture of a house and the model would give a very low value like 0.001 telling you it doesn't recognise many features in common.

Ideally this model would evolve in a Long short-term memory way meaning that after you input new kind of images, returning to the example mentioned before, after thousand of images of houses processed it would again return a low value if inputed a cat or dog image.

Answer 1

The buzzword you are looking for is anomaly detection using neural networks. Usually how this is done is that you train an autoencoder on your training data and when a new image come you calculate the difference between the original image and the image reconstructed by the autoencoder. If your autoencoder was trained on furry animals, than the reconstruction of another furry animal will have smaller error than a reconstruction of a house. Of course, there are more ways to do it