I don't personally work in this area of programming, but it's been interesting to follow how this new library was designed. It takes advantage of Elixir's Lisp-style macros to create a...
I don't personally work in this area of programming, but it's been interesting to follow how this new library was designed. It takes advantage of Elixir's Lisp-style macros to create a domain-specific language tailored to numerical computation.
The new defn (numerical definition) construct outwardly looks like a normal function def, but under the hood it gets compiled to a computation graph. This allows for automatic differentiation as well as further compilation to highly specialized code tailored to the type/shape of data the function operates on.
Here's a video of the library co-author (and creator of Elixir itself) explaining the library and using it to live-code a neural network for the MNIST dataset.
I'd be interested to hear thoughts on this from anyone with more experience in the data science field!
I don't personally work in this area of programming, but it's been interesting to follow how this new library was designed. It takes advantage of Elixir's Lisp-style macros to create a domain-specific language tailored to numerical computation.
The new
defn(numerical definition) construct outwardly looks like a normal function def, but under the hood it gets compiled to a computation graph. This allows for automatic differentiation as well as further compilation to highly specialized code tailored to the type/shape of data the function operates on.Here's a video of the library co-author (and creator of Elixir itself) explaining the library and using it to live-code a neural network for the MNIST dataset.
I'd be interested to hear thoughts on this from anyone with more experience in the data science field!