I have a question about the use of calibration in neural networks and its relation to accuracy. In Guo et al. it says: "Because the parameter T does not change the maximum of the softmax function, the class prediction remains unchanged. In other words, temperature scaling does not affect the model’s accuracy."
This makes sense to me, but then I don't understand why it is used as a method for uncertainty estimation. Suppose we use a softmax function to determine our probabilities, which is known to be bad to measure uncertainty. Now, if we use temperature scaling or something similar for calibration, do the probabilities calculated with the softmax function become true confidence measures? If this were the case, I don't see why this would be necessary, since the accuracy of the model should not be affected by the calibration.
I think I misunderstood something in the procedure and would be very glad if someone could help me!