Synapse Product Development的动态

Transcript

In a previous blog post, we developed a project called Procyon which used machine vision to process a prerecorded video just to demonstrate how we might be able to use machine vision to aid in the process of sorting recyclable materials. Later we expanded upon that project to put it into a live demo that we could show to clients coming into the office or demonstrate at tradeshows, but that project ran on a bulky gaming PC's. We decided to embark on a journey to see if we could get it to run on this. Which is a tensor processing unit about the size of a USB thumb drive. Let me show you how it works. Now, as I was saying, the previous version of this project ran on a gaming PC, which is really cool because it has a graphics card in it. And the graphics cards are neat because they can do multiple concurrent processes at once, unlike normal processors which can only do linear progressions of computations. Graphics processors, on the other hand, were designed to do 2 dimensions because they're meant for PC gaming, and the pixels on your screen are laid out in an array. 2D array and these processors are intended to make computations for all of those simultaneously. That was cool about tensor processing units like this one is they're really good for neural networks and machine vision specifically because the dimensions of the arrays or the the tensor networks that need to be processed aren't just two dimensions or three dimensions. So these the output of these is measured in Tera operations per second. So as long as the array or matrix fits within that bound. You can do one computation on the entire neural node all at once. So we went out there and downloaded multiple data sets in order to retrain a neural network to detect bottles and then reduce the size of that network and it to be able to run on one of these tensor processing units. You can see a recording of this model running live on a tensor processing unit here, and as we move different types of bottles in and out of the frame, it annotates their position within the image for you to see. Now, to demonstrate how these GPUs work, we miniaturized a previous version of a project that we've kind of already thought of, but there are many different applications out there that you could envision using machine vision on an embedded core. For example, I've got a friend who woke up in the middle of night to some noise in her kitchen and she walked in there and found a raccoon. The raccoon came in through their pet door and maybe got himself a little snack out of the dog food bowl and it wasn't really clear who was more scared, the raccoon, the dogs or my friend. But things carried out of there and everything was hunky Dory. You could envision adding some sort of locking mechanism to a pet door and use of small embedded processor to detect. You can't face or your dog's face in front of a camera and only unlock that pet door if it was your pet out there rather than some other hungry critter. And again, there are many things you can envision using these low power tensor processing units for things like reducing waste like with Procyon or maybe even Wildlife Conservation, or perhaps just improving our day-to-day lives. What do you think? Drop some examples down in the comments below and maybe we can come up with some really cool, really inventive stuff.