Implementing geometric mean for consensus opinion/levels_mean

[ryan-caesar-ramos]

Hi, I'm trying to implement the consensus opinion (levels_mean) as a geometric mean of the top-down predictions, bottom-up predictions, attention-weighted average of same-level embeddings, and embeddings of the previous time step as described by the original paper. Any ideas on how the weights should be set?

At first I thought this could be a learnable parameter, but section 9.1 reads

For interpreting a static image with no temporal context, the weights used for this weighted geometric mean need to change during the iterations that occur after a new fixation.

which leads me to believe that these might need to be outputted on the fly a la vanilla attention as opposed to being learned. Maybe an MLP that takes in the four source embeddings and outputs four scalars as weights?