fg6hr

10 ms·

61.
by fg6hr 7y ago
Let's consider a simple image classifier: 64x64 grayscale input, 10x1 output that detects 10 classes of images. The model: y=tanh(Ax+b). You'd probably say that there is no way this will work because this model is too simple. But
62.
by fg6hr 7y ago
I believe Tesla has competent lawyers who would advice to sneak the binding arbitration clause in the purchase contract.
63.
by fg6hr 7y ago
That's kinda expected. Once a car becomes a pile of buggy software, someone needs to maintain it and fix bugs. That someone is called a SW engineer and happily charges 200-300/hour. Add on top of that the service dep fees that can
64.
by fg6hr 7y ago
Actually, I read the batch norm paper and maybe I forgot important details, but it roughly went like this: "here we add a term `b` to make sure the mean values of Ax+b are zero and that will help us with convergence; ah, and here is a
65.
by fg6hr 7y ago
It's your job, though. You can always set up 1:1s and ask from time to time "is there something you are unhappy about?" or "do you think your share of our company adequately represents your contributions?" The reali
66.
by fg6hr 7y ago
If you want to play politics, covertly talk to the founders, as only they are decision makers, find a common language with them, but keep in mind that your team lead is already doing so and is trying hard to present your achievements as his
67.
by fg6hr 7y ago
Distributed systems and ML, I would say. This is if you want to land a solid job, assuming you have the right mindset. Some mentioned here video games, but that industry is far from lucrative. There are a lot of interesting things to learn
68.
by fg6hr 7y ago
A genuine question: is there any math behind ML at all? For example, is there any solid theory, with proven theorems, that would tell us what happens when we add another conv layer here or use a 3x3 conv kernel instead of a 2x2 one over the
69.
by fg6hr 7y ago
Nothing surprising here. This is why all manufacturing has moved to China. SW engineers in the US get so much not because they are so smart, but because (1) the company have to pay premium to keep the IP private and (2) they can work on-sit