Current PhD student working in ML. Looking for a place that does a great job pushing ML into high-impact production. I've done two internships. One was a late-stage startup with lots of prediction problems, but everyone just used regression. No process at the time to integrate serious real-time ML with products, so hard to find cool problems to work on. The second was research at Intel which is a *chill* place to do ML SW because it's impossible to get fired and you have plenty of freedom, but tough to see any long-term impact. Which places are a happy middle ground? For big places (FB/Google/Amazon/etc) it would be good to know teams/areas where the impact is highest. Not too concerned about hours/pressure for now. It doesn't have to be DeepMind and WaveNet->Pixel, just something better than regression. If it helps, my background is Bayesian inference/approximate inference for neural networks. TC: 55K++, so basically 110K
Probably Google. And definitely not amazon, things are highly product focused regardless of what the recruiters or managers might say.
FB
For Product ML in FB, probably any of the Ads/ Feed/ MarketPlace/Instagram/watch - Ranking / Integrity teams (You can do the combinations). As someone posted earlier in the comments, it would mostly Comprise tweaking the parameters and ship the best models. Infra is well built. However, no theory or any technical depth. For Research ML, you got to go to FAIR. Active research streams include vision, text and speech. Speech is doing good though. So, probably low - medium impact but very high technical depth. You would be working with super smart researchers. I hope this helps. I am not sure what core data science does. It would be somewhere in the middle of two. I wont recommend that if you are aspiring for high impact or high research
How's the interview process for product ML position?
Should be straightforward. 2 coding + 1 system design + 1 or 2 ML Designs
To the first order, there are two types of ML work 1. Product ML - high impact but no technical depth. Run some hyperparmeter searches, push out model, see if it helps metrics. You may be use NN instead of regression, but your day-to-day remains the same - use a package, twiddle some parameters. 2. Research ML: low impact but high technical depth. You've seen both of the above. Middle ground is very rare. Pick your poison.
I agree with your first order characterization. I'm looking for some higher order characterizations that differentiate between places and/or teams. For example Core Data Science seems like it might be a middle ground between Data Science and FAIR.
Spartacuss are you in ML @ FB?