For triplet encoding we use a one to many feed forward network for the time delta and add sum this with the code and value token to get a triplet token. We can try other temporal position methods.
We can do absolute (time after first observation or after birth -- i.e. age) or relative (time from the previous observation) -- the curent implementation.
We can try a one to many feed forward network--the current implementation--or we can try a positional encodings using sine and cosines (maybe from here)
For triplet encoding we use a one to many feed forward network for the time delta and add sum this with the code and value token to get a triplet token. We can try other temporal position methods.