The autoscaling ignores the outliers.
e.g. when starting training loss is high (10) but quickly does (<1) if I plot the scalar I will not be able to see the changes in loss too well because the graph is on a large range (0-10)
If I ignore the outliers I will get a scale of 0-1