In the context of ANNs, is there multi-task learning iff the network has more than 1 output? This is a terminology question: in the context of artificial neural networks,
does multi-task learning occur iff the network has more than 1 output?
 A: Having multiple outputs is necessary but not sufficient for multi-task learning.
One task could be a multi-label classification task (e.g. is the input a picture of a human, a cat or a dog?) with one output node per target class. So a network with three outputs is performing a single task.
A second related task could be to predict what time of day the input photo was taken (e.g. morning, afternoon or night). So a multi-task network here could have six outputs defining the two tasks.
A: The multiple tasks in multi-task learning may not be explicitly distinct classes, but differing distributions of the same labels. For example, the following could all be considered multi-task learning frameworks:




Framework
Input
Output
Single task interpretation
Multi-task interpretation




Multi-class classification
Image
Cat, dog, fish
What animal is this?
Is this a cat?Is this a dog?Is this a fish?


Binary classification
Text
Spam, not spam
Is this email spam?
Is this English email spam? Is this Spanish email spam?Is this Russian email spam?




Or more explicitly, the multiple tasks may be different labels on the same underlying data (multi-label classification). Explicitly one can build a multiheaded model, where the backbone is shared amongst related (or indeed unrelated) tasks, and the heads correspond to the different tasks e.g:




Framework
Shared backbone
Input
Output
Task




Multi-headed
"High level" facial features
Image
Head 1: binary classificationHead 2: multi-class classificationTask 3: regression
Task 1: Is this a man or a woman? Task 2: What ethnicity is this person? Task 3: How old is this person?



