8
for better convergence. The choice of optimizer was based on the experiments carried
out while training our custom CNN architecture and theory [15].
4
Evaluation
The performance of the age estimation algorithms is evaluated based on the closeness
of the predicted value to the actual value. The metrics widely used for the age estimation
as a regression task is the mean absolute error or MAE which captures the average
magnitude of error in a set of predictions. MAE calculates the absolute error between
actual age and predicted age as defined by the equation (1).
𝑀𝐴𝐸 =
1
𝑛
∑
|𝑦
𝑗
− 𝑦
𝑗
̂ |
𝑛
𝑗=1
(1)
Where n is the number of testing samples,
𝑦
𝑗
denotes the ground truth age and
𝑦
𝑗
̂ is the
predicted age of the j-th sample.
For classification tasks (age and gender), the evaluation metric used was accuracy
which denotes the fraction of correctly classified samples over the total number of sam-
ples.