On a small neural-network-like basis, it is possible to write a very precise kind of figured bass. The computer itself can improvise upon it, using a very limited number of input parameters (5 parameters which can be further compressed as a 16-bit byte for transmission).
Usefully, in this scheme, each note is not a point but a vector. And each chord is not pinned down, but shown to be one of a limited number of locations. Choosing one of these allows improvisation in the surrounding mode.
Also usefully: melody and harmony have the same representation.