The MSE loss is like x^2 where x is the residual. We apply threshold on this residual (x) and convert it to a linear loss (which is a function of |x|) instead when the residual is greater than the threshold.
If we set threshold to less than 1, it would mean that the linear loss would be greater than the squared loss (x^2<x for x<1). We ended up applying a larger loss to small residuals.
Why we are using this expression particularly delta/2 (|y_actual-y_pred|-delta/2)?
Can't we use simply |y_actual-y_pred| for delta >= some specified value determined from absolute residual plot?
The MSE loss is like x^2 where x is the residual. We apply threshold on this residual (x) and convert it to a linear loss (which is a function of |x|) instead when the residual is greater than the threshold.
If we set threshold to less than 1, it would mean that the linear loss would be greater than the squared loss (x^2<x for x<1). We ended up applying a larger loss to small residuals.