L2 Regularization and Batch Norm

I looked at it, and it is very good: good baselines, explanations, visualizations and going deeper than a typical "it is a black box but if you copy & paste it will work".

I was surprised by the nice network vis (and I did dive into the subject before: https://medium.com/inbrowserai/simple-diagrams-of-convoluted...). The only thing that looks clunky is the text logs for training (a shameless plug: https://github.com/stared/livelossplot).

Could it be that Python is being used for rapid prototyping but OCaml for production? I've heard of several trading firms with this approach.

They've also used OCaml for deep learning: https://blog.janestreet.com/deep-learning-experiments-in-oca...

Although it seems to be an OCaml binding into TF, rather than a native implementation.

There is a project that aims to bring NN/ML/DL/RL (along with other scientific computations) in the OCaml world - Owl[1]. They also have a list[2] of the potentially interesting ideas for new contributors to take.

[1] http://ocaml.xyz/

[2] http://ocaml.xyz/project/proposal.html

This is a popularization of things already published in open papers, so it does not reveal anything specific about their activities. Any place employing deep ML practitioners could have written this.

It could even be a red herring, as the most popular application of batch norm is to Deep CNNs, and those are mostly used on computer vision problems. CV does not seem important for option pricing, which is AFAIK Jane Street's big money maker. Of course I can be very wrong about this. People have tried image data as auxiliary inputs to financial data. Or you can apply Deep CNNs to 1D data like timeseries - see WaveNet applied to timeseries forecasting.

Have you got a reference for katoGo ?

edit: nevermind https://blog.janestreet.com/accelerating-self-play-learning-...