4 ms·
Ah... The ML scientist's "Metrics are harmless. People doing stupid things with metrics on the other...". Or if I'm being too subtle, this is the ML practitone
by salawat 5y ago
Ah... The ML scientist's "Metrics are harmless. People doing stupid things with metrics on the other...".
Or if I'm being too subtle, this is the ML practitoner's "Guns don't kill people. People kill people."
Which is absolutely right, I might add; but also a stop point and consideration that should be had before you go training something on sketchy data, and unleashing it on an unknowing populace.
- elcomet 5y agoThat's not what the article is talking about. They are saying that data biases are a problem but not the only one: other things can amplify biases such as the model's architecture. So when working to remove biases, fixing the data might not be enough.
- athrowaway3z 5y agoWhat salawat is probably saying is: There would be less issue with bias (for society), if the expectation/sales pitch was different. As an analogy: Somebody sells the idea of "The Unbreakable Rope" and it breaks after you hang 10kg on it. A blog post "Moving beyond material" goes in depth on how to process material. This is a good thing for people producing rope. However, for a layman who still sees people assuming "The Unbreakable Rope" is attainable, the content is going to be a little underwhelming.