François CholletVerified account

you should be using cross entropy. Probability distributions are not an Euclidean space.

for this problem for sure, but for the random Redis application of discovering data from users or alike?

What are your targets?

then you ask the network given this user, what AD to show...

looks like it would sometimes be a classification problem and sometime a regression problem. You need support for multiple losses.

in theory Redis could guess after a few tries in most cases? If the “teachers” outputs always look like all 0s but a single 1…

black box ML is a very difficult problem. One size fits all is impossible, you need to constrain the problem upstream.

Things you need to take into account: 1) number of samples available 2) type of problem 3) class imbalance 4) overfitting ...etc