Axon: A library for creating neural networks in Elixir | Hacker News Reader

Axon: A library for creating neural networks in Elixir

github.com

158 pointssmn14 comments

I haven't had time to get very far into the author's Generic Algorithms book yet, but what I did read definitely exceeded my expectations.

https://pragprog.com/titles/smgaelixir/genetic-algorithms-in...

Really looking forward to where this goes with the Erlang 24 JIT and Elixir Nx. It's amazing all the work that's been going into performance.

The companion blog post is at https://seanmoriarity.com/2021/04/08/axon-deep-learning-in-e...

If you have not been following the work José and Sean have been doing lately, I recommend you watching José’s talk introducing Elixir Nx https://youtu.be/fPKMmJpAGWc

This is probably an incredibly dumb question, but is there a good motivating example why I would want to train a neural network with Elixir? Inference, I totally get, for online systems that is amazing. But I’m having a hard time (perhaps due to a lack of imagination) thinking of scenarios where I would use Elixir to train and tune instead of the existing, fortified systems.

It's a fair question. I don't plan on using this, but I can see the value of reaching for this if your tech stack is already in Elixir.

basically all purposes will be implemented in a general purpose language over time.

if you are hiring and training people to work on elixir/erlang systems you can have them work on this as well.

performance? that's a guess of course.

What I mean is that given the rapid pace of development of deep learning, it seems hard to imagine a programming language preference will overrule the desire to have access to the latest implementation of new algorithms, or implementations that are battle hardened by wide deployments. Perhaps there's some niche this could open up though that would incentivize people to try to keep up.

Pretty incredible that you can do this type of work on the BEAM now. Wow!

Does this mean Sean Moriarity's next book will be on neural networks? :)

Maybe ;)

Please take my money. :) If you're reading this and want to learn genetic algorithms in Elixir, Sean's book is excellent: https://pragprog.com/titles/smgaelixir/genetic-algorithms-in...

So if I want to execute an existing model (e.g. produced using tensorflow/tensorflow lite) from within elixir, what's the easiest way currently?

We’re working on a way to import and export models from other formats just by reading the graph into our model representation.

""" The wait is over, my friends !

The mechanical brain is about to speak !

The mechanical brain will speak !

The mechanical brain is `* (MatchError) no match of right hand side value`

"""

Damn, there must be a typo somewhere.

Reply on news.ycombinator.com