eric-tramel / ample Goto Github PK

There should be a choice to switch the location of the learning updates (i.e. prior parameter updates and the noise learning update) to occur after the convergence of the AMP iteration. This would be more in line with EM.

When using a full, explicit, matrix, ample gives correct performance in the following cases:

• Known noise and known prior parameters.
• Known noise and unknown prior parameters.
• Unknown noise and known prior parameters.
• Unknown noise and unknown prior parameters.

However, when a mean-approximation is used for the F2(...) and F2T(...) operations, delta learning breaks down, working in the following cases:

• Known noise and known prior parameters.
• Known noise and unknown prior parameters.
• Unknown noise and known prior parameters.
• Unknown noise and unknown prior parameters.

This may relate to the manner in which Delta is being updated.