Haskell is actually really good for that. Hasktorch[1] is high quality, batteries included, and leverages Haskell's GADTs for some cute stuff like gradual tensor shape checking. There's also just a grain to Haskell that feels really good for the machine learning domain. It doesn't just alleviate whole classes of bugs and a lot of annoying background noise reasoning, but it's also a much more natural-feeling expression.

The "gotcha" is that Haskell is heavy duty machinery, and getting up to speed with it if your background lacks solid type-theory can be really daunting. For that reason alone, it could never be the default. Sometimes I like to think about how much of a disservice academia has done to itself by training mathematicians without giving them the foundational knowledge they need to utilize the nuclear-grade tooling they themselves have the most potential to benefit from. For a number theorist? Sure makes sense. But the fact that machine learning courses don't have rigorous undergraduate prerequisites in learning the foundation of computation is pure absurdity.

[1] - http://hasktorch.org/