The first versions of Gorgonia[0] (obvs not called Gorgonia) was written in Haskell. I found it difficult because Accelerate wasn't there yet when I wrote it.

The first versions of Gorgonia[0] (obvs not called Gorgonia) was written in Haskell. I found it difficult because Accelerate wasn't there yet when I wrote it. And cofounders found it difficult to understand (because Haskell was a strange language). On the one hand it is easy to implement a computation graph in Haskell. On the other hand it is not so easy to implement an efficient kernel for which the computation of the values would run - you end up writing weird looking C that looks like Haskell, which I submit is not necessarily a good thing. Working with Gorgonia IMO has been the clearest thing for deep learning for me. Granted it's a bit biased because I wrote the damn thing. Nonetheless, writing it first in Haskell had its benefits, the structure was clear and hence it has imparted some design decisions in the Go library. From time to time I still have to add new features to Gorgonia, so I still plot out roughly using Haskell. Nowadays I recommend Grenade for anyone wanting to play with deep learning in Haskell [0] Gorgonia is a TensorFlow/PyTorch equivalent written in Go. https://github.com/gorgonia/gorgonia https://github.com/gorgonia/gorgonia if you are interested.