Under what conditions should different integration increments be used?

[mars0i]

e.g. should sigma-dt sometimes be different from phi-dt? Should phi-dt sometimes differ for different levels?

On p. 7 col 2, end of section 3, Bogacz says:

"Thus on each trial we need to modify the model parameters a little bit (rather than until minimum of free energy is reached as was the case for phi)."

i.e., I think, this means that updating sigma and theta should be done more gradually, i.e. over more timesteps, i.e. based on more sensory input filtering up, than updating phi. What about epsilon? That should go at the speed of phi?

cf. end of section 5, where he says that the Hebbian Sigma update methods (which I'm not using, initially) introduced there depend on phi changing more slowly. But isn't that the opposite of what I just said??

Also, should the higher levels also go more slowly?? i.e. as you go higher, you update less often? After all, they are trying to develop generalizations that reflect a wider range of data. You don't want them changing radically due to a single outlier. Or will it all come out in the wash, and a smaller dt isn't needed?

[mars0i]

See docs/notes/example_1.md for an answer that seems to be working.