Gadgetory

hello there i'm gary sims from andrew authority now we've seen some phones release recently that have kind of neural processing hardware built into the phone you think about the ki-rin 917 which is in the main 10 and you think about some of the new i fold with their neural engines there seems to be this new wave of devices that have Hardware dedicated to doing machine learning so the question before us a day is this do you need to have dedicated hardware and NPU inside of your smartphone to experience the benefits of machine learning well let me explain so let's start with the basics what is machine learning while learning of course is the process that by experience you modify your behavior now when we come to the idea of machine learning really today we're talking about recognition pattern recognition voice recognition image recognition and so on and so when we're talking about machine learning we're really thinking about the fact that Google assistant can understand the word we're saying we're thinking about that if you show a picture to a smartphone you can tell you that it's a dog or a cat or a horse or a landscape and that can be used for things like changing the photo settings accordingly now there is a difference between artificial intelligence and machine learning machine learning as I say is that part about recognition whereas the important part of artificial intelligence is the word intelligence which of course includes concepts like abstract thinking and planning and of course then we get into a whole bunch of other areas like freewill and you know this kind of stuff that really is a whole separate idea now I do have two videos on this channel one that talks about machine learning in detail how it works I have another one that talks about artificial intelligence in detail and what that might mean for the future of mankind so please do go and check out those two videos now when we're talking about machine learning today we're basically looking at Earth's smart phones are doing recognition of things like our voices but most of that recognition happens in the cloud and what that means I speak to Google system I ask it a question and there is some pre-processing that goes on on the phone but then the results of that pre-processing go up to the cloud where the final bit of processing is done and then also the response comes if I say you know will it rain tomorrow basically those words are kind of digitized they're kind of divided up on the phone already and then they go up onto Google's cloud where it will then work out what that question meant and then try to reply with the relevant information now computing kind of goes in cycles now way back like 50 years ago and maybe even 20 years ago there was really this phenomenon that all computing power was centralized and everything at the remote side was dumb so way back you had kind of mainframes and mini-computers and they were just dumb terminals or even dumb workstations but all the work was done on a central computer and then we had kind of the PC era which brought all that computing power back to the desk so now what I wanted to do things I didn't rely on a central computer it was happening on my own private machine and now since the beginning of the internet and the cloud and mobile computing we've kind of got this hybrid situation where a lot of things I do on my mobile device whether there be a laptop or whether that be a smartphone but a lot of it is also done in the cloud and as I said machine learning at the moment is really a cloud-based kind of technology now as with we went from dumb terminals to kind of pcs also now we're coming to the stage where we need to bring some of that machine learning back down to our local devices not realized so heavily on the cloud now to do that you're going to need some processing power in your actual device whether that is a laptop or a smartphone or a Chromebook or something like that now there are several different ways of doing this at the heart of all this machine learning you need to run a neural network now I won't go into the details of a neural network now but basically the idea is is that a signal is received into a neural network and then it gives a result and you train the neural network by presenting it with a set of input whether it be an image would there be voice or would there be of something else and then you say this is a particular thing and it makes the neural network wreck nice it strengthens the connections between the neurons so that when it then sees a similar but not exactly the same data it can still produce the same output so if I say the word yellow I have my British accent and I say it in a particular way but I can also understand American accents which is why I can watch films from Hollywood and likewise Americans can watch British films Doctor Who and Downton Abbey and so on and you still understand our accent and that's because our brains are capable of recognizing a whole vast range of different sounds and voices and accents now a computer has to be taught to do those things and at the moment it's trained and that information stays mainly in the cloud now to run a neural network on your phone there are three or four different ways that it can be done it can be run on the CPU absolutely it can it can also be run on the GPU because GPUs are good at doing neural network processing it can be run on something like a DSP or what we would traditionally call a DSP and of course it can be run on these new neural processing units now when Huawei showed their curve the NPU working what they actually did will show a demonstration of how their NPU was faster at doing image processing compared to a GPU and a cpu but what it also showed us is that a CPU and GPU can do these tasks it's just that it's faster in the new NPU so really we are now in the age of heterogeneous computing which means when you have a task the computer needs to decide which part of its hardware is actually best at doing that task now for example arm have their compute library which is a library designed specifically for doing machine learning tasks and depending on the platform that's or not depending on the hardware that it has around it it will perform exactly the same functions but it will give it either to the CPU or to the GPU or to wherever it thinks it can do that tasks the most efficiently and Qualcomm have a similar idea with their hexagon DSP if you use the SDK specifically for Qualcomm processors then you can do a lot of tasks specifically in the the dias DSP and that'll also be true for Huawei when they want develop to building specifically for their keyring 970 chipset there will be an SDK that will say here is how you pass over those tasks to the NP you so in this cloud-based world that we live in at the moment you don't particularly need any special hardware in your phone to do machine learning however the point is is we would like to see live machine learning which is where the phone is actually learning not not using a preset of trained images a preset of trained words that are happening up there in the cloud but it's happening locally and a good example of that would be security for example if you want to add extra security using a neural network it's actually quite a good idea arm recently released their security manifesto which talks about how we need security in an end-to-end solution from the device all the way through to the cloud and they were talking about how a neural network is one way to achieve that so the phone can learn your patterns it can learn where you do most of your shopping it can learn where you normally carry your phone in your pocket you can learn at what speed you walk at and then it can make decisions about whether someone a star on your phone and run off with it or whether someone's trying to use your phone to make a payment in a place that you don't normally go to things like that and that requires machine learning that happens locally on the device and that's where things like an NP you or a neural networking engine come into play but today where things stand of course it's just about speed so we know that the while wane Kieran 970 is much faster maybe three times faster at doing image recognition when it uses its own dedicated hardware we know that Apple have included the neural engine so it can do the face ID processing and that's an example of where it's learning live because it's a different face for every owner so we need to learn that live and we know the pixel two has got this image processing unit in it we don't know much about it yet but again that dedicated hardware that we'll be doing something to do with machine learning so what does all this mean for the future what it means is we really are in the infancy of this dedicated hardware sort of push I remember way back when I got my very first 3d graphics card and plugged it into my PC and I started to play and it completely changed everything but that that way of doing things now how that graphics can achieve that is no longer even a possibility today we've moved on so we moved on from the innovators to a kind of a standard way of doing things and that's really where we are now today with new processing units we have the innovators now that are trying these things out but I can guarantee you that in three years time the way we do this we're very different to the way we're doing it today and the key thing here is machinery is really gonna become interesting when we start doing live learning which means the phone is able to use this new processing hardware to learn something new today not just relying on training that's received elsewhere the problem with that though is that with bad training you get bad results so it's really gonna be interesting to see how OMS and the scientists and the researchers come around to using NPU hardware in our devices or as I said will we just keep on using the CPU and GPU because this problem of live training is actually too hard to overcome well my name's Gary Sims from Andhra thority I do hope you enjoyed this video if you did please do give it a thumbs up also please don't forget to subscribe to Angela college's YouTube channel please follow us on social media and last but not least do go over to Andrea Orta comm because we are your source for all things Android