The paper – which is pretty short – goes into all of the details. But punchlines: 1) it works. With minimal impact on training, majority vote of samples gets pretty close to deterministic accuracy in classification. 6/

2) More interestingly, the 2nd most popular choice reflects uncertainty you would naively expect.  Shirts and jackets on Fashion MNIST; dogs and deer on CIFAR-10.  1’s and 7’s on MNIST. 3) This amazingly seems pretty robust to precision in the probabilities. 7/