Fizz Buzz in Tensorflow (opens in new tab)

This whole thing is making fun of asking fizz-buzz to a senior programmer, but yet I've found it to be one of the best phone screen questions possible.

For someone who is truly a senior programmer, they knock it out in about 30 seconds and we move on. For the ones who pretend to be senior programmers on their resume, it trips them up and I know right away that their resume is either a pack of lies or their previous coworkers were helping them a lot more than they let on.

If I had actually had this person, I would have probably laughed as soon as he did the imports, said, "clearly you think this is a silly question" and then explained why I ask that question. Hopefully we could have moved on from there.

interviewer: How far are you intending to take this? me: Oh, just two layers deep

haha

This is the most off the wall and hilarious thing I've seen in a month. For calibration: I spent the whole weekend at Maker Faire.

No wonder you didn't get the job, using a shallow ann when they were probably looking for a pointer network.

I can't tell if this guy did this during a real interview, or if that's part of the joke. Maybe that's intentional. Very funny, either way.

It reminds of the paper where they use SVM to "visually identify" the matrix rank:

http://www.oneweirdkerneltrick.com/rank.pdf

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.

Not that I advocate this approach, but another way to look at FizzBuzz is as a personality test.

If a (senior) developer balks at writing FizzBuzz, perhaps it shows arrogance and a lack of humility. It could also be useful to see if the developer follows up to inquire as to why they were asked a FizzBuzz question, and about its relevance to the work. I see this kind of questioning as not only fair game, but good to ask about. Not asking about something that seems out of place could seem off.

Maybe the one-hot binary encoding wasn't the best feature set. Base 3 and Base 5 encodings might have worked better.

Please tell me you really did this. A company name would make my year, but I'll be satisfied with a simple yes/no confirmation.

It's a fun article, but solving the problem with the simplest, clearest solution possible should be the goal and if someone provided this solution I would hesitate to hire them. Yak shaving in training a neural network would actually make me nervous that the candidate would over-engineer solutions. It's not like whiteboarding sleep sort or something for a fast easy clever solution (if impractical and inefficient). The guy who has to maintain this version of fizzbuzz after it's wrote would loose his mind. I'm probably being overly practical - it's a very fun article.

I think I'd start with, "Let's define two combinators S and K..." ....

Fizz Buzz in combinatory logic, that would be quite entertaining.

I'm absolutely delighted that there is more and more such developer-generated prose with context of recent events. Is there a place where it would be aggregated?

Seems like a bit of a wasted opportunity to assume input as binary digits, rather than as a sequence of hand-written numbers on the white-board. Then there should of course be a second whiteboard, and a robot arm with a marker for output, along with a pair of cameras for stereo input and feedback while training the robot arm in hand-writing "fizz", "buzz" and "fizzbuzz" ...

Should've trained to generate the output strings directly, using an LSTM decoder... That would be more end-to-end.

I think this just illustrates the Alpha principle, the interviewer was obviously intimated. This is because they are a Beta interviewing and Alpha. If the interviewer had more imagination and confidence rather than the long pauses it could have been a great learning opportunity!

   >>> for i in range(1,101): print "FizzBuzz"[i*i%3*4:8--i**4%5] or i

The above line cost your interview :) He might be expecting like this one

All that and it still couldn't get the right answer? With a basic pattern like that?

I suspect somewhere Hofstadter is having a good laugh.

Great post. Both funny and informative.

And here I thought that the Haskell style FizzBuzz was clever.

It would be much more interesting to see a TensorFlow program actually learn to solve FizzBuzz from examples, instead of hardcoding in the logic:

    if   i % 15 == 0: return np.array([0, 0, 0, 1])
    elif i % 5  == 0: return np.array([0, 0, 1, 0])
    elif i % 3  == 0: return np.array([0, 1, 0, 0])
    else:             return np.array([1, 0, 0, 0])