Neural Networks In CPUs

[ldo17]

A few months ago, Samsung announced its Exynos M1 processor, which sports a neural network in its branch-prediction system. Now, we hear about AMD’s new “Ryzen” (formerly “Zen”) processors, which also use a neural network for something similar.

Neural networks seem like magical things: they can be trained to perform recognition tasks that we don’t know how to program. Trouble is, once the neural network has been trained up, we have no idea how it works. Also, they can sometimes behave peculiarly when encountering situations that were not covered in their training.

To me, the use of such things is almost like an admission of defeat: the problem is too complex to solve with our own brains, so let us summon the demons (or genies, if you prefer) to solve the problem for us...

[Zan Lynx]

With enough effort at analysis, neural networks can be figured out. But as more and more nodes are added, the number of complex interactions becomes too difficult.

This is not unique to neural networks and machine learning however. There is plenty of spaghetti code written by humans that is impossible to figure out. That's why big areas of computer science are dedicated to reducing the number of interaction points in software, and reduce "cyclomatic complexity." Modular code, object oriented code, data hiding, function hiding, all of this is not to make code writing easier. It is to make understanding and maintenance easier.

But anyway, your last paragraph is a very good point. I don't see it as an admission of defeat though, just a good use of time. Instead of attempting for a poor, slow human to understand computer code and teach a computer to solve the problem, that human creates a method where the computer itself learns how to solve the problem. And instead of creating one solution and making it fit all the possible code in the world, the computer's neural net continuously updates and responds to the code that's running right now.