4 ms·

The original paper (http://www.oxfordmartin.ox.ac.uk/downloads/academic/future-of-employment.pdf http://www.oxfordmartin.ox.ac.uk/downloads/academic/future-o...

by panic 9y ago

The original paper (http://www.oxfordmartin.ox.ac.uk/downloads/academic/future-of-employment.pdf http://www.oxfordmartin.ox.ac.uk/downloads/academic/future-o...) derives these numbers in the following way: 1) A selection of 70 occupations is hand-labeled as "automatable" or "not automatable" by the authors and a group of ML researchers. 2) Three statistical models are trained using nine variables (Finger Dexterity, Manual Dexterity, Cramped Work Space/Awkward Positions, Originality, Fine Arts, Social Perceptiveness, Negotiation, Persuasion, and Assisting and Caring for Others) taken from O * NET (https://www.onetonline.org https://www.onetonline.org) survey data for each of 702 occupations. The models are graded by how well they match the hand labels from step 1. The best model of the three is chosen. 3) The numbers generated by this model are then reported for all 702 occupations. This includes the ones that were labeled to begin with: "We implicitly assumed that our hand label, y, is a noise-corrupted version of the unknown true label, z. Our motivation is that our hand-labels of computerisability must necessarily be treated as such noisy measurements. We thus acknowledge that it is by no means certain that a job is computerisable given our labelling." This methodology doesn't make a lot of sense to me. If you allow yourself to make up (or "hand-label") 70 numbers, why not make up all 702? Why not trust your own labels in the final data? What is the statistical model even telling you when it's being trained on labels you don't trust?

avaer 9y ago
I find it sociologically interesting how layers of math and publishing indirection can be used to convince people that someone's opinions and biases (hand labels) might in fact be statistically significant facts about the world. To their credit, at least the original authors acknowledge this problem.
- crdoconnor 9y ago
  It's even more interesting to consider why economists might have these inherent biases. What incentives might they be responding to?
in9 9y ago
> This methodology doesn't make a lot of sense to me. If you allow yourself to make up (or "hand-label") 70 numbers, why not make up all 702? Well, what is the science in that?? See, the headline will tell you that the X% of chance that Y% job will be automated, and we know we have to take that with a grain of salt. However, the statistical models have some extra information, such as which of the 9 variables carries more information about how to predict the labels and etc. Also, some model can also handle somewhat "bad" data, since even though it is crappy, some signal is still there. All that said, we still need to take the prediction with a grain of salt... :D

by panic 9y ago

avaer 9y ago

I find it sociologically interesting how layers of math and publishing indirection can be used to convince people that someone's opinions and biases (hand labels) might in fact be statistically significant facts about the world. To their credit, at least the original authors acknowledge this problem.

crdoconnor 9y ago
It's even more interesting to consider why economists might have these inherent biases. What incentives might they be responding to?

in9 9y ago

> This methodology doesn't make a lot of sense to me. If you allow yourself to make up (or "hand-label") 70 numbers, why not make up all 702? Well, what is the science in that?? See, the headline will tell you that the X% of chance that Y% job will be automated, and we know we have to take that with a grain of salt. However, the statistical models have some extra information, such as which of the 9 variables carries more information about how to predict the labels and etc. Also, some model can also handle somewhat "bad" data, since even though it is crappy, some signal is still there. All that said, we still need to take the prediction with a grain of salt... :D