custom labels

• Sometimes we need to change our target labels to make it work better with our loss function
  • for example, we may use a custom form of label smoothing
• How to improve your labels
  • if your labels has the form: $\sum \frac{t}{r}$, consider using $\sum \frac{t^2}{r^2}$
    • play with how this looks:
      • https://www.desmos.com/calculator/gdehmkoyeq
    • the distribution of the labels looks like an inverse bell graph
      • there are more values with smaller labels, and more with larger labels