Ma-Lab-Berkeley
Uncertainty Estimation
Hi,
Very impressed by your work. And I have been wondering since the ReduNet is a white-box, one should be able to write down what is the uncertainty of the ReduNet's prediction analytically. Say in the test phase, I feed an image of half apple half orange to the ReduNet (which is trained to classify apple and orange), I should be able to get the prediction uncertainty for free? And in theory, I should also be able to track back through every layer to see how the uncertainty propagate, right? Is uncertainty estimation in your roadmap?
I believe yes, that you should be able to compute the projection between your feature vector from test data and subspace. You should be able to project at every step.
I don't know too much about uncertainty estimation but hopefully there are ways to expand our framework into other research topics. Thank you for the suggestion!
