World Model
If you haven’t read my previous blog post about the “world model” please go check it out.
Training the world model is a prototypical example of Self-Supervised Learning; Learning the mutual dependencies between its inputs. It is said that a general formulation can be done with a framework of Energy-Based Models(a function producing low energy values when x and y are compatible and higher values when they are not). I also have a blog post on Energy-Based Models, so if you are interested please go check that out as well.
Adding a Latent Variable
To evaluate the degree of compatibility between x and y, an EBM may need to consider a latent variable. By adding a latent variable, we can parameterize information that represents y which x alone does not contain.
If we find the latent(z) that minimizes the energy E(x,y,z), the resulting Energy F(x,y) only depends on x and y.
Reference: Research Paper
