Given the uncertainty distributions are non-Gaussian, it does not make sense to use a standard deviation of internal network inferences to determine the error bars on network predictions. An alternative non-parametric, robust, computationally light yet accurate solution is required. A straightforward approach that fulfils these criteria is to calculate the [2, 16, 50, 84, 98] weighted percentiles of the inferencer and forecastor predictions every epoch. These percentiles are easy to interpret in relation to the more usual standard deviation as the median, +/-1 sigma, and +/-2 sigma values of the distribution.
1 Like
Yes, fully agreed! This is a great place to start with. We can work on refining this in the near future.