I am trying to create a multivariate regression model in keras, but the model always ends up predicting a single value regardless if the input.
I tried fixing the learning rate, batch size and model architecture, but so far I can’t fix it.
Upon further inspection, I know that some values in the output vector will be hard or impossible for the model to predict (say the height of a man deduced from the car brand he likes), but due to the nature of the problem, I don’t know which these hard or impossible to predict variables.
I am currently using MSE loss. I suspect I have to use or create a custom loss function that deals with this kind of problem.
Chatgpt proposed a Heteroscedastic loss, but I can’t understand what it means.