State importance of favoring algorithms

[janEbert]

Lesson 5, question 1:

Which of the following would you apply Deep Learning to?
[...]
2. Calculating the median and interquartile range of a dataset.

Answer:

2. no, even though it is a regression problem, you'd need tons of training data and still be off by quite a bit

The answer misses one of the most important points: there are existing, super efficient and 100 %-accurate algorithms for this. :)

[psteinb]

Ok, I guess I have to reformulate this. What I mean was to try to learn arithmetic operations using Deep Learning.

[janEbert]

I'd say the point still stands – there are already algorithms for arithmetic operations which don't require

• the whole bother of training a neural network, and
• an expensive neural network evaluation that
• only returns approximate results.

[psteinb]

I think there is a misunderstanding here ... that is the sole point, that learners discover that quite some tasks (like arithmetic operations) do not make sense to solve with a neural network. ;-)

[janEbert]

Yes, but the answer only explains that the problem may be hard with a neural network. It doesn't mention that it's simply not the right tool for the job. :D

That's what I'm getting at; the question is great to get people thinking, but the answer smudges the more important lesson (the task being one that ML should not be applied to) by arguing that neural network training is hard.
At least, that's the impression I get when reading the answer. Could also be that you want the students to think about this and come to the conclusion on their own; in that case, disregard everything I said. :)

[psteinb]

I take that point. Could you please send a PR to fix that? ;-)