Setting aside the question what "machine learning" actually means... Machine learning is just software and if it is a computer running software, then machine learning is already "built in" to some extent. I assume that the claim is that there is some specific hardware support to accelerate an algorithm currently in vogue for machine learning. Can someone comment on technically what kind of hardware acceleration is considered state of the art for machine learning and if this is what on the table for the Raspberry Pi? What kind of practical speed improvement would this hardware in the Pi create? Would it double, triple, or raise by an order of magnitude the speed that a model can be trained or executed? In other words, is this actually significant or not? The article doesn't provide any useful context.
The quick eli5 type explanation would be, largely parallel floating point math. Which is why GPU's are so heavily used currently.
The main difference is in a lot of cases high precision isn't required, so 16-bit floating point math is used instead of 32-bit to get far higher performance with similar memory bandwidths.
Can you eli8 and also explain the differences between say a classical GPU and what the "Tensor cores" in NVIDIA terminology do? I'm a complete luddite when it comes to this but it seems at the moment for machine learning the only GPUs people are recommending for any machine learning stuff are the RTX ones from NVIDIA, and if it were only floating point math would simply a higher core count be sufficient rather than special purpose silicon?
the differences between say a classical GPU and what the "Tensor cores" in NVIDIA terminology do?
Truly Classical GPUs had fixed function hardware designed to do specific graphics calls. You feed it a 3D model and a texture map for color and tell it where your camera is and it returns an image. You couldn't really program anything to run on a GPU you could only fill in the blanks. You mad libs the 3D scene, it returns an image.
Modern GPUs have moved to programmable rendering. They are fundamentally now no different from a CPU but have made different performance choices in how they're laid out. A CPU
while you could use vector instruction sets for some of it, like normalization operations. there are other functions that aren't in old SSE that are neccessary to accellerate a convolutional neural network.
Yes at the core of things, "built in" machine learning is acceleration of floating point vector math...
But to say you are including built in machine learning, also implies there would basically be libraries around this were you could pass in common neural network models already trained, and be able to pass input through them.
There's a reason why the camera is also part of one initial package, as support will probably be heavily focused around neural networks that deal with images (for things like face recog
"Yes at the core of things, "built in" machine learning is acceleration of floating point vector math..."
That's not wrong; it's not insightful either. Literally anyone could trivially deduce this, but no one on/. appears able to go beyond it either. What a surprise. Where are all those programming experts?
ML doesn't merely require vector operations, it requires execution of specific kinds of operations on a massive scale. Specifically, it requires huge numbers of convolution operations, but it's OK, Su
Can someone comment on technically what kind of hardware acceleration is considered state of the art for machine learning
The new features described in TFA are designed to speed up neural networks using platforms such as TensorFlow. You want fast operations, especially multiplication, on large matrices of low-precision floating-point values, typically FP16.
Training a NN is very computationally expensive, so even with the new features, it is not something you will want to do on a RPi. It makes much more sense to upload your data to the cloud, do the training there using TPUs or high-end GPUs, and then download the fully traine
Maybe if you weren't so obviously prejudiced and convinced it was nonsense and bothered learning something about machine learning you wouldn't be so cynical or even *gasp* might actually find a use for it.
You read like the old man dispariging the internet. "My telephone works just fine. I've never needed to connect to no InterNet or whatever fad these kids are wasting their lunch money on these days! I don't see no reason to add dedicated networking hardware... Probably just dials phone numbers faster or so
i dont have an average of 15 phd's (not even a hi-skool rag) so im still stuck on "how is a.i. different from google collecting your clicks" ? a glorified database with a search algorithm forming clusters on most hits - despite that it seems to be able to compose music by itself . . . (but so do pop-tarts)
owh - aHA , the/. pitfall... not falling for it today
What exactly is "built in" machine learning? (Score:3)
Re: What exactly is "built in" machine learning? (Score:4, Interesting)
The quick eli5 type explanation would be, largely parallel floating point math. Which is why GPU's are so heavily used currently.
The main difference is in a lot of cases high precision isn't required, so 16-bit floating point math is used instead of 32-bit to get far higher performance with similar memory bandwidths.
Re: (Score:2)
Can you eli8 and also explain the differences between say a classical GPU and what the "Tensor cores" in NVIDIA terminology do? I'm a complete luddite when it comes to this but it seems at the moment for machine learning the only GPUs people are recommending for any machine learning stuff are the RTX ones from NVIDIA, and if it were only floating point math would simply a higher core count be sufficient rather than special purpose silicon?
Re: (Score:2)
the differences between say a classical GPU and what the "Tensor cores" in NVIDIA terminology do?
Truly Classical GPUs had fixed function hardware designed to do specific graphics calls. You feed it a 3D model and a texture map for color and tell it where your camera is and it returns an image. You couldn't really program anything to run on a GPU you could only fill in the blanks. You mad libs the 3D scene, it returns an image.
Modern GPUs have moved to programmable rendering. They are fundamentally now no different from a CPU but have made different performance choices in how they're laid out. A CPU
Re: (Score:2)
Wow thanks for the effort and the writeup.
Re: (Score:0)
I think "machine learning" is just today's trendy application for vector instructions like you'd get with MMX and Altivec.
Re: What exactly is "built in" machine learning? (Score:2)
while you could use vector instruction sets for some of it, like normalization operations. there are other functions that aren't in old SSE that are neccessary to accellerate a convolutional neural network.
What it likely means (Score:1)
Yes at the core of things, "built in" machine learning is acceleration of floating point vector math...
But to say you are including built in machine learning, also implies there would basically be libraries around this were you could pass in common neural network models already trained, and be able to pass input through them.
There's a reason why the camera is also part of one initial package, as support will probably be heavily focused around neural networks that deal with images (for things like face recog
Re: (Score:2)
"Yes at the core of things, "built in" machine learning is acceleration of floating point vector math..."
That's not wrong; it's not insightful either. Literally anyone could trivially deduce this, but no one on /. appears able to go beyond it either. What a surprise. Where are all those programming experts?
ML doesn't merely require vector operations, it requires execution of specific kinds of operations on a massive scale. Specifically, it requires huge numbers of convolution operations, but it's OK, Su
Re: (Score:2)
Re: (Score:3)
Can someone comment on technically what kind of hardware acceleration is considered state of the art for machine learning
The new features described in TFA are designed to speed up neural networks using platforms such as TensorFlow. You want fast operations, especially multiplication, on large matrices of low-precision floating-point values, typically FP16.
Training a NN is very computationally expensive, so even with the new features, it is not something you will want to do on a RPi. It makes much more sense to upload your data to the cloud, do the training there using TPUs or high-end GPUs, and then download the fully traine
Re: What exactly is "built in" machine learning? (Score:2)
Maybe if you weren't so obviously prejudiced and convinced it was nonsense and bothered learning something about machine learning you wouldn't be so cynical or even *gasp* might actually find a use for it.
You read like the old man dispariging the internet. "My telephone works just fine. I've never needed to connect to no InterNet or whatever fad these kids are wasting their lunch money on these days! I don't see no reason to add dedicated networking hardware... Probably just dials phone numbers faster or so
Re: (Score:1)
?
did it learn something ?
Re: (Score:1)
owh - aHA , the
Re: (Score:2)
repeat if(rnd(3) !=0)?good++:bad++; until(good===100||bad===100) ? did it learn something ?
Two out of three ain't bad?
Re: (Score:1)