Can I define n+1 outputs for n class classification?

[bornexmachina]

I am sitting on an anomaly detection task and I had the following idea.

Assume I have 3 classes and each of them could have an anomalous sample. Then, I would train with 4 classes, the last one reserved for the anomaly. But in my training data I would never actually have any data on the class 4.

Now during inference, I would try to classify the data and if it is an anomaly it should have low scores for class 1 to 3 and thus be labeled as class 4.

Would this approach even compile? So expecting 4 class weights but having labels only for 3 classes?

[ptrblck]

@bornexmachina

Now during inference, I would try to classify the data and if it is an anomaly it should have low scores for class 1 to 3 and thus be labeled as class 4.

This might be a good idea but I don’t think the model will automatically classify every “unknown” as class4 assuming it was never trained to output this label.

@bornexmachina

Would this approach even compile?

Yes, it would technically work since you don’t need to use all logit outputs.

[bornexmachina]

technically correct is the best form of correct. Thanks!