6 ms·
You're implicitly making the assumption that an algorithm is unbiased and less likely to be fallible - and unlikely to be susceptible to bias or manipulation (a
by cnewey 9y ago
You're implicitly making the assumption that an algorithm is unbiased and less likely to be fallible - and unlikely to be susceptible to bias or manipulation (at least, compared to a human). There is a considerable amount of academic and popular literature showing that these assumptions are false, at least at a technical level.
The problem of vetting YouTube videos isn't necessarily something that can be solved by an algorithm, at the end of the day. Lots of the simple cases, sure - but such an algorithm won't work on even the most moderately challenging content. After all, moderating complex content such as video doesn't just require a well-trained neural network to recognise objects portrayed in images and sounds in audio... deciding whether a video is appropriate or not for monetisation is something that requires an explicit understanding of the social and cultural context which the video is going to be viewed in - which is both highly subjective and difficult to express in an algorithm.
NB;
- Neural networks (and other AI techniques) can be tricked and manipulated (see [0,1]; examples of adversarial images)
- Machine learning algorithms can induce or confirm bias through inadequate training data and poor assumptions. (see [2]; a book about bias in machine learning algorithms)
[0]: https://arxiv.org/abs/1510.05328 https://arxiv.org/abs/1510.05328
[1]: http://www.evolvingai.org/files/DNNsEasilyFooled_cvpr15.pdf http://www.evolvingai.org/files/DNNsEasilyFooled_cvpr15.pdf
[2]: https://weaponsofmathdestructionbook.com https://weaponsofmathdestructionbook.com
- notyourday 9y ago> You're implicitly making the assumption that an algorithm is unbiased and less likely to be fallible - and unlikely to be susceptible to bias or manipulation (at least, compared to a human). There is a considerable amount of academic and popular literature showing that these assumptions are false, at least at a technical level. No, I'm arguing that algorithms are transparent, which means that they can be adjusted. > The problem of vetting YouTube videos isn't necessarily something that can be solved by an algorithm, at the end of the day. Lots of the simple cases, sure - but such an algorithm won't work on even the most moderately challenging content. of course they would here: while (!decision_pass_the_spot_check()) { tweak_algorithm() } > deciding whether a video is appropriate or not for monetisation is something that requires an explicit understanding of the social and cultural context which the video is going to be viewed in - which is both highly subjective and difficult to express in an algorithm. Certain cultures think it is acceptable to stone women. In their societal context it is OK. We do not, however, care that it is OK in their societal context. So we make sure that in our algorithm stoning women ==> demonetized. Now we move onto the next problem. Did someone flag a video that we processed that had women stoned that was not demonetized? We open a ticket with the people responsible for that model, with the video that was not demonetized and have them figure out why it did not happen. That's how you solve a complex problem. You break them into simpler ones and when you see an issue you address it. Claiming that the problem is just too complex for algorithms is avoidance.
- cnewey 9y agoYour argument that algorithms are transparent is equally flawed. There is an entire field of research dedicated to introspection of layers in convolutional neural networks! A system that is complex and nuanced enough to be able to deal with a problem such as auto-moderating YouTube videos will be phenomenally difficult to inspect (and nearly impossible to "adjust" in the way that you seem to be suggesting). To clarify; I am not saying that it will never be possible to solve this sort of problem with an algorithm - just that doing so would require solving the small inconvenience of general artificial intelligence first... :-). It certainly isn't beyond the realm of possibility in the (very distant) future, but to suggest that it's possible right now (and that it would be better than a human!) is to over-egg the pudding somewhat.
- notyourday 9y agoIt is not flawed at all. If you cannot provide a validation for your algorithm decisions when challenged then you algorithm belongs in a research lab or on your laptop, not in production.
- joshuamorton 9y ago>It is not flawed at all. If you cannot provide a validation for your algorithm decisions when challenged then you algorithm belongs in a research lab or on your laptop, not in production. Alright, then you cannot use an algorithm for image or video recognition. So how do you solve this problem?
- majewsky 9y agoOkay, then your argument is valid... in a different layer of reality. Meanwhile, everyone and their dog are rolling out blackbox neural net classifiers to production
- notyourday 9y agoAnd i cannot wait until one of these companies ends up in court trying to explain why its engineers cannot explain how a decision was made.