basic

Last updated Feb 27, 2024 Edit Source

#work/patientsim 2024-02-07 13:59

I should really go with the transformer on this one. The triplets do turn the data into a sequence of values, some with much higher relative importance than others. That feels pretty attention-native. The idea is much more incremental, too. The original triplet paper simply uses a self-supervised prediction objective, and I could add only the VAE into the mix.

I mean, look at this graphic: triplet-transformer-G7LK8EX5 Looks a whole lot like Keith’s formulation of one of our core problems!