Manager: Okay, we need a system for detecting pedestrians in real time on the road!
Engineer: I’m sure there’s an R package that does it! Alternatively, we can use Scikit-SelfDrivingCar or something. I bet we can get it done the next week!
A few hours of googling later…
Engineer: Apparently, there’s nothing like that. Let’s do some heavy lifting - we will use Scikit-VisionMagic, download a dataset for it in .csv format from Kaggle, call a bunch of standard methods from tutorial and ship it to production!
Next day:
Engineer: Uuuh, there’s no readily available datasets, no out-of-the-box libraries…Sigh. Alright, let’s use Keras and someone’s project from Github!
A few denigrating comments from StackOverflow later…
Engineer: Okay, I’ve managed to run it on our data, but this model gives us some weird artifacts that were not reported in the instructions…Guess, I’ll have to learn Keras deeper and try to fix that.
A few questions on Quora later:
Manager: Hey, how’s that pedestrian thing going?
Engineer: Pretty good, just a few more minor fixes!
muffed sound of crash in the background
Engineer: Turns out this model was good only for demonstration! I’ll have to write my own in TensorFlow…
A few days of copy-pasting tutorial code later:
Manager: Dude, we are in no rush, but you’ve told us that you’d be done by now.
Engineer: Yeah, there were some complications, need to make sure everything goes smoothly after the release.
Do you hear that sound? That’s the sound a gradient makes when it explodes [1]
Engineer: This pre-implemented loss function is a mess! And I will have to write a couple of my custom layers, and a loss function, and the regularization that’s not shipped in the available library…so many things to study.
One ML course later.
Engineer: Turns out the receptive field was too narrow…if only I took a class on deep learning…
Engineer: Oh, that batch normalization thing really rocks! Why don’t they write it in scikit-learn tutorial?
One Goodfellow’s book later.
Engineer: Finally! I’ve broken a ton of things on my own and made a ton of mistakes that could be avoided, but hey, that’s the fun of learning!
Engineer: Alright, it works, the simulations look pretty good!
Deployment engineer: Great! But your system requires 24Gb of memory and can’t process more than 2 frames per second on a high-end Titan card. We can’t use it in a car.
Engineer: Damn it! It looks like I’ll need a fast inference method and reduced memory consumption…if only I knew the computational complexity of every part of my network and the amount of parameters like they teach in the algorithms course…And what’s about that FP16 thing anyway?..
Manager: We have run out of funding and everyone is fired.
The moral of the story is that you can’t outgoogle good fundamental education and relevant technical knowledge.
Follow me on :-
Youtube : https://goo.gl/2exULj
Instagram(Tech) : https://goo.gl/ikLTyb
Instagram(Food) : https://goo.gl/ht3bZM
Quora : https://goo.gl/sEoxY3
Medium : https://goo.gl/M3QvUq
Linekdin : https://goo.gl/LeaGSF
Steemit : https://goo.gl/o7RUj7
GitHub : https://goo.gl/3R38LK
Zomato : https://goo.gl/tMKpM5
