Different error metrics have been experimented with:
Mean Square Error (MSE)
The quadratic error in absolute values is the same for values close to one but decreases faster in approach to zero. This means that the gradient’s reaction to significant errors is stronger, but it corrects minor inaccuracies less effectively.
Model loss:
MAE Loss + Maximum Deviation
An opposing approach amplifies the reaction to errors. In addition to the averaged error, the worst discrepancy is also considered. This should direct the gradient against averaging. Despite being just one pixel, convolution is a repetitive operation where one parameter affects multiple pixels. The maximum deviation should be aimed at the convolution, which excessively averages values.
Model loss:
MSE Loss + Maximum Deviation
A hybrid approach. Such a combination may either offset each other’s drawbacks or simply hinder each other’s operation.
Model loss:
With a small number of iterations, no significant results were observed in all three cases.
Author — Egor Zyryanov
Leave a Reply