mean_squared_distance_to_scaled_one_hot

probly.quantification.mean_squared_distance_to_scaled_one_hot = <flextype.singledispatch.Flexdispatch object>

Mean over members of \(\| h_k - s\, e_{\arg\max_j h_{k,j}} \|_2^2\); class axis is last.

Parameters:
  • sample – Logits sample across ensemble members.

  • scale – Scale s; None uses K (num classes). Matches DARE Eq. 35 [dMDMV23].