Well, AI can "imagine" the full contents of complex 3D scene simply by being shown three images at various angles, so I thought there might be a possibility to

Well, AI can "imagine" the full contents of complex 3D scene simply by being shown three images at various angles, so I thought there might be a possibility to train an AI how to look for patters in crypto and plaintext and "imagine" the function as well. Any way you can point me to a paper (or otherwise) which explains why this is mathematically impossible?