I am faced with a problem which I bet was already solved before, but that I had never seen. Perhaps by discussing it abstractly, someone can point me to relevant literature.
It goes like this: I have a dataset of images $I_j$ and numerical features $f_{1,j}, f_{2,j}, ..., f_{k, j}$. In production, I don't have access to the features $f_{i,j}$, which I know to be more relevant to my learning problem than the images.
However, theses features $f_{i,j}$ are also very hard to obtain, and there is a known correlation (informal) between the images and the numerical features $f_{i,j}$. So the whole point of using the images is to avoid using the $f's$.
Know comes the question: is there any useful way I could learn a representation of the images that "exposes" the maximum information about the features $f_{i,j}$? Perhaps some form of autoencoder with a regularization term that penalizes uncorrelated codings?
Any ideas are welcome. Thanks!
Edit:
In the problem that I have, the images $I_j$ are $(64,64)$ images of tumor nuclei, and the $f_{i,j}$ are genetical information about each cell, such as the expressivity of certain biomarks (HER2, ER, etc.).
