3 ms·

Article 25 is interesting. I don't know if there is a similar right under GDPR or the California data protection law. Article 25: When using personal informati

Article 25 is interesting. I don't know if there is a similar right under GDPR or the California data protection law. Article 25: When using personal information to conduct automated decision making, the transparency of the decision making and the fairness and reasonability of the handling result shall be guaranteed. Where an individual believes automated decision making creates a major influence on their rights and interests, they have the right to require personal information handlers to explain the matter, and they have the right to refuse that personal information handlers make decisions solely through automated decision making methods...

cblconfederate 6y ago
It's part of GDPR too. Though probably impossible to implement since people simply dont know why ML systems make their decisions.
- datameta 6y ago
  The thing is even many ML researchers don't believe that interpretable models are possible without losing accuracy. There's a good episode of Data Skeptic [0] with one of the co-authors of a paper [1] delving into this. They talk about real-world ML models used to predict recidivism as well as medical expert systems. [0] "Black Boxes Are Not Required" https://podcastaddict.com/episode/107514072 https://podcastaddict.com/episode/107514072 [1] "Why Are We Using Black Box Models in AI When We Don’t Need To? A Lesson From An Explainable AI Competition" https://hdsr.mitpress.mit.edu/pub/f9kuryi8/release/5 https://hdsr.mitpress.mit.edu/pub/f9kuryi8/release/5
  - twblalock 6y ago
    Black box models are going to be a problem when we let them make life and death decisions like driving cars. At some point, one of these models will do something wrong, like hit pedestrians on a regular basis, and a judge will order its owners to "make it stop doing that". Or maybe a new law will be passed to the same effect. Then the owners of the ML model will be in a bit of a pickle if they don't actually know how it works.
    - Cerium 6y ago
      One way I have seen the integration of black box models in situations where incorrect decisions are unacceptable (in this case a medical environment) is to implement a good enough solution to the problem using conventional programming techniques, and then give the black box model free reign to provide more accurate answers bounded by the conventional system. In the case I saw, the black box model was able to improve the accuracy by about 10x over the existing system for 90% of the use cases. The other 10% fall back to the existing system and do no worse.
      - candiodari 6y ago
        I don't get this because we've seen decision making under pressure, for judges. The conclusions are invariably that even the most reliable indicators are violated occasionally (human intelligence is of course a multi-armed-bandit algorithm so this is exactly what you'd expect, and want). Whether the people making decisions agreed they were experimenting or not did not matter at all. They also weren't able to explain this (and even when they did, they just provided the explanation one of the lawyers provided, which sounds fine, until you realize that of course two lawyers provided an explanation. So they are giving exactly the wrong information to understand the decision: judge chooses between "A, because ..." and "B, because ...", and at that point your decision is about the difference between A and B, not about A or B themselves) Whilst I have not seen a comparable study with doctors' decision making ... I have very little doubt it's exactly the same. Doctor's AREN'T reliable and are in fact, no matter how hard they deny it, are experimenting on patients and learning from it. Of course, when you think about more than a single patient, this is what you want. With this I mean you want (ideally many small, some large) mistakes to happen as long as it isn't you under the knife, but you're expecting at some future time it'll be you, because that way the doctors will maximally improve by then. Once you're under the knife, I guess you don't want to see experiments, but yeah, there's a price for skill. And there's a price to be paid in mistakes for human skill as well. Even for doctors. Even for surgeons.
    - cblconfederate 6y ago
      Pharmaceuticals are much more of a "black box" . It's all about statistics, and if tesla seriously wants people to use SD, it will have to be a lot better, beyond placebo statistical significance.
detaro 6y ago
GDPR has Article 22 and Article 13 2f. https://gdpr-info.eu/art-22-gdpr/ https://gdpr-info.eu/art-22-gdpr/ https://gdpr-info.eu/art-13-gdpr/ https://gdpr-info.eu/art-13-gdpr/

cblconfederate 6y ago

It's part of GDPR too. Though probably impossible to implement since people simply dont know why ML systems make their decisions.

datameta 6y ago
The thing is even many ML researchers don't believe that interpretable models are possible without losing accuracy. There's a good episode of Data Skeptic [0] with one of the co-authors of a paper [1] delving into this. They talk about real-world ML models used to predict recidivism as well as medical expert systems. [0] "Black Boxes Are Not Required" https://podcastaddict.com/episode/107514072 https://podcastaddict.com/episode/107514072 [1] "Why Are We Using Black Box Models in AI When We Don’t Need To? A Lesson From An Explainable AI Competition" https://hdsr.mitpress.mit.edu/pub/f9kuryi8/release/5 https://hdsr.mitpress.mit.edu/pub/f9kuryi8/release/5
- twblalock 6y ago
  Black box models are going to be a problem when we let them make life and death decisions like driving cars. At some point, one of these models will do something wrong, like hit pedestrians on a regular basis, and a judge will order its owners to "make it stop doing that". Or maybe a new law will be passed to the same effect. Then the owners of the ML model will be in a bit of a pickle if they don't actually know how it works.
  - Cerium 6y ago
    One way I have seen the integration of black box models in situations where incorrect decisions are unacceptable (in this case a medical environment) is to implement a good enough solution to the problem using conventional programming techniques, and then give the black box model free reign to provide more accurate answers bounded by the conventional system. In the case I saw, the black box model was able to improve the accuracy by about 10x over the existing system for 90% of the use cases. The other 10% fall back to the existing system and do no worse.
    - candiodari 6y ago
      I don't get this because we've seen decision making under pressure, for judges. The conclusions are invariably that even the most reliable indicators are violated occasionally (human intelligence is of course a multi-armed-bandit algorithm so this is exactly what you'd expect, and want). Whether the people making decisions agreed they were experimenting or not did not matter at all. They also weren't able to explain this (and even when they did, they just provided the explanation one of the lawyers provided, which sounds fine, until you realize that of course two lawyers provided an explanation. So they are giving exactly the wrong information to understand the decision: judge chooses between "A, because ..." and "B, because ...", and at that point your decision is about the difference between A and B, not about A or B themselves) Whilst I have not seen a comparable study with doctors' decision making ... I have very little doubt it's exactly the same. Doctor's AREN'T reliable and are in fact, no matter how hard they deny it, are experimenting on patients and learning from it. Of course, when you think about more than a single patient, this is what you want. With this I mean you want (ideally many small, some large) mistakes to happen as long as it isn't you under the knife, but you're expecting at some future time it'll be you, because that way the doctors will maximally improve by then. Once you're under the knife, I guess you don't want to see experiments, but yeah, there's a price for skill. And there's a price to be paid in mistakes for human skill as well. Even for doctors. Even for surgeons.
  - cblconfederate 6y ago
    Pharmaceuticals are much more of a "black box" . It's all about statistics, and if tesla seriously wants people to use SD, it will have to be a lot better, beyond placebo statistical significance.

detaro 6y ago

GDPR has Article 22 and Article 13 2f. https://gdpr-info.eu/art-22-gdpr/ https://gdpr-info.eu/art-22-gdpr/ https://gdpr-info.eu/art-13-gdpr/ https://gdpr-info.eu/art-13-gdpr/