You can find the difference between recurrent and recursive neural nets on wikipedia. No need to edit your post then flag someone when they point out how lazy you are in understanding a simple rant post, pointing out how lazy people are in understanding the "AI" fads.
I assume this article is directed towards people new to neural nets who think they mastered AI. However, in neural network research noone cares about of the things the author is making fun of. These are examples from tutorials beginners learn in the first week, not what experts are remotely concerned with.
This is a pretty stupid article. No one is pretending that they're a genius, but this guy is just full of inaccuracies. Let me break it down to save other people time and energy (I have a little to spare):
1. "but that complexity comes from repetition and a random number generator" - Most of the complexity comes from trying to solve non-convex optimization problems via SGD. "Repetition" doesn't make sense unless you're talking about vectorization (you're not), and an rng has little relevance.
2. "Congrats! You took the above code, and looped the loop again." - sure, if that were true then people would have trained deep neural networks in the 90s. Instead we needed researchers (see? geniuses) to invent batch normalization, highway networks, CNNs, dropout, etc. Also, that's not an "recursive neural network", that's a recurrent one.
3. "So you trained a neural network using Nvidia GPUs and moved it to the phone…" - this entire section makes so little sense I don't know where to start. Neural networks are, for the most part, robust to small rng differences - that's why they have surprising generalizability. I have no idea what this guy is trying to say about phones and GPUs though.
4. "What it does well is help you visualize what is happening in those 11 lines" - well, that's not the point, but I'm starting to understand that this guy doesn't get it.
5. "Building a neural network with 1 trait for every word in the English language would require a network that used as much computing power as all of Google." - I'm just going to let that one sit there.
...
"There is neural network code in my tool box."; sure, you sound like you've imported Tensorflow before (possibly?). But maybe before dismissing an entire field you should figure out how any of it works.
I guess I agree with the article in that simply using an existing neural network, treating it like a blackbox (which to me it is), and feeding it cropped photos pulled from an API is not rocket science. Also I see a lot of 'data scientists' portfolios where they basically just took existing ML projects and tweaked them slightly, wrote a few lines of python, and tried to make themselves seem like cutting edge devs. But yeah I respect that there are true geniuses behind the development of these techniques themselves, and even just training a blackbox NN can be tough.
This is a terrible article. He makes a total straw man out of neural networks by pulling code from an "Intro to Neural Networks in Python" tutorial and saying "What's so hard about this?" He also doesn't explain it correctly, not even close. Maybe he should actually do the tutorial.
What's the point of this? That he's as smart as people who build AI? This person, Brandon Wirtz, needs to work on his self-confidence, not just intelligence.
Anyone who is learning something new is a genius in my opinion - especially if they are excited about it and want to share their results so that others can learn.
10 comments
[ 1.9 ms ] story [ 29.1 ms ] threadAlso it's a post from another blog... I'd wander over there and try to figure it out, but lunch beckons...
1. "but that complexity comes from repetition and a random number generator" - Most of the complexity comes from trying to solve non-convex optimization problems via SGD. "Repetition" doesn't make sense unless you're talking about vectorization (you're not), and an rng has little relevance.
2. "Congrats! You took the above code, and looped the loop again." - sure, if that were true then people would have trained deep neural networks in the 90s. Instead we needed researchers (see? geniuses) to invent batch normalization, highway networks, CNNs, dropout, etc. Also, that's not an "recursive neural network", that's a recurrent one.
3. "So you trained a neural network using Nvidia GPUs and moved it to the phone…" - this entire section makes so little sense I don't know where to start. Neural networks are, for the most part, robust to small rng differences - that's why they have surprising generalizability. I have no idea what this guy is trying to say about phones and GPUs though.
4. "What it does well is help you visualize what is happening in those 11 lines" - well, that's not the point, but I'm starting to understand that this guy doesn't get it.
5. "Building a neural network with 1 trait for every word in the English language would require a network that used as much computing power as all of Google." - I'm just going to let that one sit there.
...
"There is neural network code in my tool box."; sure, you sound like you've imported Tensorflow before (possibly?). But maybe before dismissing an entire field you should figure out how any of it works.
It's mastery and smart application that makes all the difference.
What's the point of this? That he's as smart as people who build AI? This person, Brandon Wirtz, needs to work on his self-confidence, not just intelligence.
This is just condescending vitriol.