about detection score

[alinlab]

about detection score

[sangwoomo]

Hi, @William-oliver,

1. If x is in-distribution with respect to the original distribution, s(x) would also be in-distribution with respect to the transformed distribution. Here, sim(s_1(x_1), s_1(x)) measures the OOD-ness of s_1(x_1) with respect to the transformed distribution. Intuitively, we ensemble the detection score over all shifting transformations.

2. Most prior work uses the softmax probability as a detection score. Eq. (8) is just a logit values (before softmax) of the auxiliary classifier, which resembles the prior work. Eq. (9) simply aims to combine two scores. One need to set a threshold, which can be computed from the training data; it trade-offs the precision-recall (a.k.a. ROC curve).

3. SimCLR objective aims to decrease the cosine similarity of z1 and z2. Here, one can decrease the absolute distance of z1 and z2, or increase the scale of both to decrease the relative scale of the absolute distance. We hypothesize that the later one is an easier way to decrease the cosine similarity.

[William-oliver]

thanks for your quick reply.