Uncertainty Estimation
At Desupervised we have put a lot of effort into developing tooling such as Borch that allows us to build models that have a proper uncertainty estimation. Uncertainty comes in very handy when working in high-stakes situations, say one wants to use image classification to determine if an x-ray scan shows signs of cancer. In that situation, you want a model that clearly shows when it is uncertain about a prediction such that you can have the scan reviewed by a doctor. As the consequences of a false negative could literally be the patient's life.
But uncertainty comes at a cost, primarily at a computational cost that makes the predictions take longer time and more complex logic may be needed to handle the uncertainty estimations correctly. The degree of slowdown depends on how precisely the uncertainty is estimated, ranging from 2x slower to 100x slower for high-stakes applications.
We build models both with and without uncertainty, mostly the decision comes down to how much value it adds to the specific use case and how important speed is. If it is unclear what is needed for your project, we will gladly give our recommendation of what we believe is best suited for the problem at hand.
We build models both with and without uncertainty, mostly the decision comes down to how much value it adds to the specific use case and how important speed is. If it is unclear what is needed for your project, we will gladly give our recommendation of what we believe is best suited for the problem at hand.