Loss-weights focus on Os, maybe at a 10:1 ratio to Bs and Is. This way, the model is focused on not falsely predicting labels to be present places it sholdn't.
Recall based loss
Loss-weights focus more on correctly finding Bs and Is, which is what has been the key to actually finding some hard metrics over zero. This loss focuses so much on guessing correct labels that it over estimates where labels should be, often predicting labels where Os should be present.
This is a detailed look into loss weighting.
Precision base loss
Loss-weights focus on
Os, maybe at a 10:1 ratio toBs andIs. This way, the model is focused on not falsely predicting labels to be present places it sholdn't.Recall based loss
Loss-weights focus more on correctly finding
Bs andIs, which is what has been the key to actually finding some hard metrics over zero. This loss focuses so much on guessing correct labels that it over estimates where labels should be, often predicting labels whereOs should be present.