I have been reading some paper on age prediction from facial image for my undergraduate thesis work. I see that most of the paper use mean absolute error as model performance metric rather accuracy. Even if the model had softmax layer in the final layer. Could you help me with that?
Given that age is a numeric measurement, please explain how one might use accuracy as a metric?
What problem is the model trying to solve when having softmax as final layer?
1 Like
In some datasets, age is represented as an integer value rather than a floating point value. In such cases, each age becomes a discrete class, and the goal is to predict which class (age) an image belongs to. However, it seems unusual to use Mean Absolute Error (MAE) as the evaluation metric when using a softmax classification layer. MAE is typically used for regression tasks, not classification. I am new to this, so I feel confused. 
I agree, this is unusual.