AUC ROC socre seems decent even before learning starts

[seunghyeon528]

Hi, thx for the code!

Actually, I came across with a question, running your soruce code.

I found that AUC ROC socre is already above 90 % even before learning starts.

I guess this is because classifying algorithm (cosine similarity of 2 nearest data) is quiete effective.

The overall model owes a lot to hard-coded classifying algorithm even rather than deep-learning.

I'm new to anomaly detection, So I'm wondering if I've understood correctly.

thank you!

[talreiss]

Hi,

In prior work (PANDA) we demonstrated that combining pre-trained features with simple anomaly detection (e.g. kNN) results with great performance. Since that our feature extractor is a large ResNet pre-trained on the ImageNet dataset, the results are pretty good from starting epoch.

[seunghyeon528]

Thx for such kind explnatation.

Hope you have a good day!

2022년 3월 6일 (일) 오전 6:29, talreiss @.***>님이 작성:

Hi,

In prior work (PANDA https://arxiv.org/pdf/2010.05903.pdf) we demonstrated that combining pre-trained features with simple anomaly detection (e.g. kNN) results with great performance. Since that our feature extractor is a large ResNet pre-trained on the ImageNet dataset, the results are pretty good from starting epoch.

— Reply to this email directly, view it on GitHub https://github.com/talreiss/Mean-Shifted-Anomaly-Detection/issues/7#issuecomment-1059833802, or unsubscribe https://github.com/notifications/unsubscribe-auth/ASOYDGACFN2DRMGRVL5JMNDU6PG2HANCNFSM5IBR5XHA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you authored the thread.Message ID: @.***>

[talreiss]

Your welcome :)