Description
Description
For historical reasons pt.min returns pt.neg(pt.max(pt.neg(x))). This seems to have been the case mostly to avoid having to define Min Op and its gradient. There is a later "uncanonicalize" phase that converts those expressions to min, suggesting we prefer them, but don't put it in place because of the lack of gradient.
We should reassess this as is adds some unwelcome complexity. The L_op implementation (cleaned up in #901) works directly for min. R_op, on the other hand uses Argmax (not sure why this is needed in the forward but not backward pass, CC @aseyboldt), so a similar Min.R_op may need to use Argmin. Similar to Min that's currently implemented as Argmax of negative of x, which is probably fine? We could also consider a direct Argmin but that is not as ubiquitous and hence less annoying.