which one is positive one sample and negative sample in your work?

[alinlab]

which one is positive one sample and negative sample in your work?

[sangwoomo]

Hi, same as SimCLR, we use the same images with different augmentations as positive samples and different images as negative samples.
However, unlike SimCLR, we think some of the augmentations produce different (or distributionally-shifted) ones from the original.
In original SimCLR, they simply ignored those augmentations (e.g., rotation) since considering them as positive samples degrade the performance.
In contrast, we found that treating them as the negative ones in addition to the original negative samples (i.e., different images) improves the OOD detection performance.

Formally, say x1 x2 are two images, T1 T2 are two augmentations (that produces positive samples), and S is a strong (or distribution shifting) transformation.

Then, the original SimCLR ignores S and think

• T1(x1), T2(x1) are positive
• T1(x1), T3(x2) are negative

and ours additionally consider S by

• in addition to the upper assumptions, T1(x1), T2(S(x1)) are also negative

Hope this clarifies your question.
Please let us know if you have any unclear points!