Dear Author,
From the article what I understood that a multi-task problem is where you devise a model that can do multiple tasks like classification and regression on similar datasets. However, the example that you have provided is a multi-label (multi-class) classification problem. Am I missing here something? Moreover, since it’s not a binary problem binary cross-entropy will provide a distorted error but why categorical cross-entropy can’t be applied here is baffling to me.
I reproduced the work using keras’s functional API and used categorical crossentropy as the loss function. The network did just fine.
https://gist.github.com/RedowanNafi/15d346da5a74db8bcf4ce0ad6a5684fb
Overall, the example, especially the inference part was enlightening. Thank you.
