CVPR Daily - Tuesday

9 DAILY CVPR Tuesday Matthias Hein " Neural networks are provably overconfident far away from the training data " " Adversarial Confidence Enhanced Training (ACET) mitigates the problem... " " ... but asymptotically the problem of overconfident predictions persists " Matthias says that the next step is to provide a modification of neural networks so that they have provably the guarantee that far away from the training data we have uniform confidence over all the classes. He adds that in general, we need provable guarantees for neural networks. This can have several flavours, but we have to ensure that what we are doing does not produce something harmful for society. Computer vision is applied in safety- critical domains, such as autonomous driving and medical applications, so it’s very important that we know what we are doing. To find out more, come along to Matthias’ oral [1.1A] at 09:23 today and to his poster [5] at 10:15-13:00.