Hacker News new | ask | show | jobs
by BenderV 3684 days ago
He should have used a Recurrent Neural Network with Long Short Term Memory so that the neural network won't have been dependent of the maximum number given NUM_DIGITS. Scalability! No wonders he didn't got the job.

Joke aside, it's funny how simple machine learning problems can reveal people who think you can just give neural networks anything and output anything, and it will work like magic.
2 comments
vintermann 3684 days ago
It also reveals something that's not obvious, but is very clear to anyone who's tried (deliberately or accidentally): Teaching a neural network simply to count is surprisingly hard.
link
Bromskloss 3684 days ago
> people who think you can just give neural networks anything and output anything, and it will work like magic.

What _is_ the proper method?

link
BenderV 3684 days ago
Well, there is a variety of possible answers here. Giving the direct integer ("1", "23") to a simple deep network won't work.

- RNN with LSTM might be a better approach, since it will more based on the value of the ordered bits than their "disposition", and could scale to an arbitrary high number.

- @zardo mentioned Pointer Network (https://arxiv.org/abs/1506.03134). It looks to solve this kind of problems but honestly I'm discovering it now.

- Give the base 3 and 5 of the input would be the fastest solution. (but since it's kind of hardcoding part of the solution, it's more a trick than a general solution).

Can anyone correct me what if I got something wrong here? Did I missed something?

link