3 ms·
2 reasons. 1. the quality of the recommendations is the real competitive advantage. It's like saying that all the sports scouting agencies need to fully reveal
by screye 6y ago
2 reasons.
1. the quality of the recommendations is the real competitive advantage. It's like saying that all the sports scouting agencies need to fully reveal their algorithm to finding top talent to ensure fairness, when that exact algorithm is the thing that makes them a top scouting agency.
2. The company itself doesn't really know how the recommender works.(source: I built explainable ML models for a living).
The inherent black box nature of any non-linear function approximator makes it near impossible to enable Explain-ability and gauge fairness for any system. In fact, I'd go as far to say that a sufficiently complex system of any kind quickly goes beyond human understanding. There are some decent proxies for it (SHAP, attention maps, gradient maps), but many recent papers raise questions on whether they really capture what we think they capture.
I think regulation of some form is certainly a good idea for tech ethics. However, it does appear that laws are being drafted by politicians, lawyers and social scientists. There is a glaring lack of qualified ML/CS researchers and practitioners on the implications of vaguely worded policy to their current processes and pipelines.
I was at ACM FAccT (ACM's fairness,accountability and transparency conference) this year and too many of the attendees seemed to be more concerned with the 'what needs to be achieved' rather than the 'how we can go about achieving it'. Terms such as fairness and accountability do not have widely accepted or principled meanings, which makes it near impossible to make a structured discussion around how anyone would go about implementing these in practice.
In conclusion, what you state is necessary, but we are very far from coming up with actually useful laws to implement it. Moving too quickly risks haphazard rules that do more damage than help.
- specialist 6y agoIt's great that you're conscientious, seriously working on ML, big data, adjacent policy issues. Thank you. In the spirit of "yes, and...": Can't manage what we can't (or won't measure. So policy wise, we should not delegate judgement, and default to hippocratic until we know more. ML & big data is following the fad cycle, just like optimization, expert systems, etc. I'm 95% certain that >95% ML and big data is hooey, most value has been realized, the rest is imagined. But meanwhile, these foolsgold pathogens are wreaking the economy. Or worse.
- screye 6y ago> Thank you I don't think I deserve it. I still work for a big conglomerate and most of the rest of my time goes into building the same algorithms that try to eek out that extra bit of accuracy. Although, we do have a lot of responsible AI, compliance, legal and interpret-ability committees we need to satisfy to make sure we aren't causing unintended consequences. > foolsgold pathogens I've flip flopped on this opinion, but over the last year or so, my opinion has solidified to being, "There is very real, massive untapped potential and value in the market, but most companies use it for marketing rather than solving the real technical problems at the core of it". I'd like to think we are the former, that solve real ML problems....but well, I'm exactly an unbiased judge of my own self. I read this really nice paper that suggested grounding ideas of AI ethics in UN fundamental human rights, rather than hand wavy ideas like accountability and transparency. Since UN Human rights are much discussed, well understood and have universal-ish acceptance, it is far easier to take policy written from that POV and actually apply it to ML/AI. I'd like to see something similar to the ACL2020 best paper "checklist", that creates a whole bunch of narrow statistical tests to validate certain traits of ML models. It'd be nice to have something like that, but for AI/ML ethics. An openly available, standardized suite of narrow tests that your ML pipeline has to pass as an audit, before being certified as "compliant" and deployed into the web. That way, the regulators get to control the test, and the companies get to control their secrets. It also ensures a level playing field and each component can be changed without an explicit dependency on the other.