Mark Koh

@RAPIDSai To sum it up: 1. Evaluation = speed 2. GPUs everywhere 3. Limit your architectures 4. Kaggle solutions are a DB 5. Refactor more 6. Pipelines > models I hope you learned something! Follow me @marktenenholtz to get more high-signal ML content!

I've spent 1000's of hours building ML models over the last couple of years. Here are some tips that would have made me work 10x faster (that you can read in 2 minutes):

I don't know who needs to hear this, but please use more #dataclasses with type hints in your #python code 🐍 docs.python.org/3/library/data…

Does anyone else do their best thinking when they're at a concert?

@arkosiorek Reinforcement Learning in recommender systems (I'm biased though haha)

what are the most interesting ML problems to work on these days?

TL;DR: Tabular: XGBoost/LightGBM/RF Time series: XGBoost/LightGBM/RF Image: ResNet/EffNet Text: RoBERTa Audio: ResNet/EffNet Your best bet is usually to start with these and then experiment from there. Nothing in ML is an end-all-be-all!

In God we trust, all others bring #data. - William Deming (theoretically)

@wilschroter My parents did this haha, they even called them "play dates"!

I'm enjoying the stage of parenting where you start to "date" other parents. The kids connect, and then the parents go on a first date and see if they have a relationship together. I don't remember my parents doing any of this...

...is it too late to start using Twitter?