5 ms·

This opacity is a huge deficit in machine learning. A seriously vast amount of work has been done attempting to address it, and machine learning algorithms that

This opacity is a huge deficit in machine learning. A seriously vast amount of work has been done attempting to address it, and machine learning algorithms that are persistently uninterpretable (neural nets being the poster-child for this problem) will always have very limited use. We understand in principle what any given neural net is doing, but our lack of understanding of the specifics makes it impossible to trust them for anything interesting. We can't tell when they might fail. For this reason a lot of people are busy solving problems that "machine learning" might one day handle, but we are doing it with relatively conventional algorithms that fall outside the "machine learning" envelope because the areas we work in (computer assisted surgery, say) aren't amenable to the application of things we don't understand. To go back to his comparison with god: the incomprehensibility of gods is generally considered a feature by religious people. The incomprehensibility of machine learning algorithms is unquestionably a bug to those of us who have hard problems that need to be solved reliably. Most of the algorithms you use in everyday life are not machine learning algorithms (and certainly not neural nets) for precisely this reason.

pixl97 12y ago
> but our lack of understanding of the specifics makes it impossible to trust them for anything interesting. We can't tell when they might fail. Sounds like we're making 'people 2.0'.
- qu4z-2 12y ago
  "Predictable, or agrees with a human implementer/trainer. Pick one." ?
ForHackernews 12y ago
It's good to point out that researchers are working to reduce this opacity, but this comment contains several unsupported assertions. > the incomprehensibility of gods is generally considered a feature by religious people. I'm not sure how you reach that conclusion. It's hard to generalize about "religious people" or even "religions" because they are so varied. I might point to the book of Job in the Christian Bible as evidence that the inscrutable nature of God (or the universe, if you prefer) is considered frustrating and tragic enough that scripture takes pains to attempt to explain it. In fact, there's an entire branch of theology [0] that tries to address this shortcoming. > Most of the algorithms you use in everyday life are not machine learning algorithms Is that true? My understanding is that almost all modern web services with vast data (Facebook, Google, Netflix) make at least some use of machine learning algorithms. [0] https://en.wikipedia.org/wiki/Theodicy https://en.wikipedia.org/wiki/Theodicy

pixl97 12y ago

> but our lack of understanding of the specifics makes it impossible to trust them for anything interesting. We can't tell when they might fail. Sounds like we're making 'people 2.0'.

qu4z-2 12y ago
"Predictable, or agrees with a human implementer/trainer. Pick one." ?

ForHackernews 12y ago

It's good to point out that researchers are working to reduce this opacity, but this comment contains several unsupported assertions. > the incomprehensibility of gods is generally considered a feature by religious people. I'm not sure how you reach that conclusion. It's hard to generalize about "religious people" or even "religions" because they are so varied. I might point to the book of Job in the Christian Bible as evidence that the inscrutable nature of God (or the universe, if you prefer) is considered frustrating and tragic enough that scripture takes pains to attempt to explain it. In fact, there's an entire branch of theology [0] that tries to address this shortcoming. > Most of the algorithms you use in everyday life are not machine learning algorithms Is that true? My understanding is that almost all modern web services with vast data (Facebook, Google, Netflix) make at least some use of machine learning algorithms. [0] https://en.wikipedia.org/wiki/Theodicy https://en.wikipedia.org/wiki/Theodicy