Here is an interesting problem:
You trained a model to classify pictures of 100 different animal species. It does a good job at it.
But when you show it a picture with a species that wasn't part of the training set, the results are obviously wrong.
How do you work around this?
You trained a model to classify pictures of 100 different animal species. It does a good job at it.
But when you show it a picture with a species that wasn't part of the training set, the results are obviously wrong.
How do you work around this?
This is also known as a "negative" class, and it helps with this problem, assuming you are capable of collecting images from unknown objects.
I've also found the advantages of this negative class to diminish as more random objects are thrown in there.
I've also found the advantages of this negative class to diminish as more random objects are thrown in there.
https://twitter.com/alephthoughts/status/1348983660054151169?s=20
It turns out that knowing what you don't know is a tough problem to solve in machine learning.
You'd expect the confidence score returned by the model to be very low for unknown objects. This is, unfortunately, not necessarily the case.
You'd expect the confidence score returned by the model to be very low for unknown objects. This is, unfortunately, not necessarily the case.
https://twitter.com/youplusai/status/1348984099369746435?s=20
Intuitively, looking at the confidence score seems like a sensible approach. In practice is less so.
It's very common for a classification model to return a very high score for unknown classes, especially when these classes share common patterns.
It's very common for a classification model to return a very high score for unknown classes, especially when these classes share common patterns.
https://twitter.com/miraculixxs/status/1348981894998876160?s=20
Remember, the problem is not whether you can improve your model. The problem is about understanding that your prediction is completely wrong even when you get a high confidence score from your model.
https://twitter.com/FahadGorain/status/1348978831881809922?s=20
Unfortunately, this is not an "overfitting" issue.
The model has never been trained with the class that we are trying to infer. However, the model still offers an incorrect prediction.
The model has never been trained with the class that we are trying to infer. However, the model still offers an incorrect prediction.
https://twitter.com/mad_enrico/status/1348980112801325057?s=20
This is a promising approach.
Notice that the heart of it is to properly detect out-of-distribution queries to our model.
Still not sure how this could be implemented in practice, but I'll dive deeper and report back any findings.
Notice that the heart of it is to properly detect out-of-distribution queries to our model.
Still not sure how this could be implemented in practice, but I'll dive deeper and report back any findings.
https://twitter.com/emirarditi/status/1348987109688926214?s=20
• • •
Missing some Tweet in this thread? You can try to
force a refresh
