Remembering the Future

In a maximally lossless compressed representation all relationships are preserved and the loss of any bit will result in the loss of information. Therefore the evolution of a representation can be analyzed in terms of lossy and noloss compression. If there is a compression to a region of a space, this can be described in terms of minimizing entropy.

You want to reduce the size of semantic voids and differentiate any aliased states. Continuity can be one kind of relationship where you constrain the number of directions a membrane diffuses. A point in model space should be close to the average of its neighbors in input space.

Topics: Representation | compression