Timeline for How to treat (label and process) edge case inputs in machine learning?
Current License: CC BY-SA 4.0
10 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Feb 5, 2021 at 23:26 | comment | added | verdery | Ideally, softmax is required to represent certainty of the model but, in reality, this is not the case. Since, the model learns to minimize error during training, it tries to assign maximum probability for its predictions. This situation causes very "confidential" model which assigns over 90% most of the time. proceedings.neurips.cc/paper/2020/file/… | |
| Feb 5, 2021 at 23:11 | history | edited | juicedatom | CC BY-SA 4.0 | added 7 characters in body |
| Feb 5, 2021 at 23:09 | comment | added | juicedatom | Yes, it is classification but you are still regressing some value between zero and one. In industry I've heard the generated labels that are used for SSD and the RCNN families as "regression targets". Which I guess could be confusing. | |
| Feb 4, 2021 at 22:37 | comment | added | nbro | Hm, I'm not sure "that represents the output of a regression target within the model." is correct, but maybe I'm misinterpreting this part. The softmax is used in the context of classification, so why are you saying "regression target"? | |
| Feb 4, 2021 at 18:57 | history | edited | juicedatom | CC BY-SA 4.0 | added 151 characters in body |
| Feb 4, 2021 at 18:53 | comment | added | juicedatom | For the softmax maybe I wasn't using the correct terminology. The softmax output depends on whatever you are regressing. In the case of SSD for example, the softmax represents the classification regression target on the class of the anchor box. | |
| Feb 4, 2021 at 18:52 | comment | added | nbro | Just one correction. You say that the softmax probability represents "specifies how certain the model is that the cat is a cat", but that's not really correct, as you even say that these deterministic models do not really model "uncertainty". See this post ai.stackexchange.com/q/24872/2444, for example. Regarding your practical note, this seems to be similar to having another class in your training dataset which is "unclear object". See also this: ai.stackexchange.com/q/4889/2444 | |
| Feb 4, 2021 at 18:45 | history | edited | nbro | CC BY-SA 4.0 | typos |
| Feb 4, 2021 at 18:32 | history | edited | juicedatom | CC BY-SA 4.0 | added 650 characters in body |
| Feb 4, 2021 at 18:24 | history | answered | juicedatom | CC BY-SA 4.0 |