Machine Learning and Effective Altruism

By mariushobbhahn @ 2021-01-16T08:46 (+11)

I have been an EA for more than 5 years now and been within the Machine Learning (ML) community for around 2 or 3. I'm currently doing a Ph.D. in probabilistic ML. in my experience, a lot of ML researchers are very open to EA ideas (essentially we get paid to make algorithms better while using less resources; this is very similar to the EA mindset) but relatively little actual connections between classical ML research and EA are made for that high level of sympathy. 

The only connection I am currently aware of is a Facebook group called Machine Learning for Good which seems to be somewhat inactive. If there is more, please let me know. 

So in the following, I want to suggest some connections between classical ML, i.e. improving decision making with algorithms when some amount of data is available (not AI safety), that seem like pretty low-hanging fruit to me. I would be interested in your opinions or whether there are good reasons this hasn't been done yet.

1. Apply ML to EA problems. A lot of organizations affiliated with EA, e.g. those rated by GiveWell are very data-driven. While GiveDirectly started using ML to more effectively predict who is poor and automated some parts of their pipeline, this seems to be the exception rather than the norm. I have no special insights into these organizations and thus can't tell whether they internally already use ML or have tried and it didn't work but, intuitively speaking, ML has the potential to yield improvements. Even if this only mean automating or impr0ving one task this could still free up resources worth a couple of 100k$ due to the size of some of these organizations.
2. Connect the ML community to EA problems. I think it's kind of ridiculous that ML researchers (me included) use pretty random benchmarks to evaluate their algorithms. Clearly, they sometimes have a specific purpose, i.e. a complexity or dimensionality, but often researchers just use what's available. So if there are any datasets where the improvement of a score would directly translate into lives saved, making them easily available and promoting them might be a very low hanging fruit.
3. Networking of EAs in ML. I started my Ph.D. when Covid began, so I haven't been to physical conferences yet but as far as I can tell there haven't been any EA in ML meetups or Workshops in the past. This might be a good place to generate further ideas and connect researchers that are sympathetic to EA. If people think this is a good idea, I would be willing to submit a workshop proposal for the upcoming ICML online conference. Since I'm in the first year of my Ph.D. it surely wouldn't hurt to have someone with more rep to support me in that effort.

These are just some first ideas and I have many more. With this post, I mostly want to gather feedback and get things moving. If you are a researcher in ML or data science and think this is a good idea, maybe comment or up vote a comment so that I get a rough impression of the general interest. If you work at an organization that has lots of data but lacks the data science / ML skills, you can also contact me or comment below and we can evaluate whether there are reasonable tools to apply or figure out the next steps.