ColdFusion - Neural Network Prototyping On the Go

Starting point is 00:00:00 You are watching Cold Fusion TV. Hi, welcome to another Cold Fusion video. Today's episode is going to be something a little bit different. Intel approached me and asked me to talk about the concept and application of neural networks to tie in with their Intel Mavidius Neural Compute Stick. First, we're going to cover what the device actually is, and then we're going to dig a bit deeper into neural networks and what they're used for. So let's begin.

Starting point is 00:00:31 So, what is the Intel Mavidius Neural Computers? stick. It's a compact device not much bigger than the standard USB drive that plugs into a computer through a USB 3 port. It enables rapid prototyping, validation and development of deep neural network interface applications and it doesn't require users to connect to the cloud. It's basically hardware designed to run certain types of AI systems super efficiently. Okay so let's back up for a minute and explore what a neural network actually is. In simple terms a neural network is an algorithm based on the principles of machine learning. So what is machine learning?

Starting point is 00:01:11 It's the ability for a computer program or AI to gradually learn by experience towards a goal established by the human programmer. For instance, if you wanted to train an AI to tell the difference between a cat and the dog, you start with the basic parameters, then you have it built, test it, then take the top performing models, rebuild and fine tune from there. This process happens over and over again, ultimately generating what we consider more intelligent algorithms after thousands of iterations. This process generates artificial neurons, otherwise known as nodes, which are connection points of data. Each of these nodes consist of an equation and have an input and an output. The nodes are arranged in layers that altogether make up the neural network. What's really fascinating about neural

Starting point is 00:01:56 networks is that there's so many of these connections and they're so elaborate that the original programmers aren't entirely sure about how they come about or what they even mean. The end result is an algorithm that can learn things, recognize patterns and make decisions in a human way and it doesn't have to be programmed to learn explicitly. It learns by itself just like a brain. And even more staggering since 2012, the general complexity of neural networks has increased 500 times. When a neural network is formed, it can do highly specialized tasks such as the previous simple example of a cat versus dog sorter. This is an application known as visual processing.

Starting point is 00:02:36 The development of generating a neural network is known as the training phase and involves vast amounts of data. After this phase, the neural network needs to be tested in real-world applications to see how it responds in real-time and if it works as desired. Traditionally, you'd need to connect to a cloud service to do this. But what if there was another way? This is where the Intel Mavidius Neurocompute stick comes in. What's fascinating about this device is that you can gain on-device deep neural network capabilities without using much power. In fact, as I mentioned before, this is considered a low-power device, which is pretty cool

Starting point is 00:03:11 because in most scenarios you'd be utilizing massive server farms. Intel's neural compute stick has a specialized on-board hardware visual processing unit, capable of more than 100 gigaflops of performance within a 1 watt power envelope. The Intel Mavidia's compute stick can calculate the vast amounts of data required in a neural network without having to connect to the cloud or even be online. This equates to an increase in privacy and reliability and it means whatever you're working on stays local and mobile. There are fewer variables inherent with working local and it also keeps things

Starting point is 00:03:44 running smoother so that you can troubleshoot quicker since you eliminate some of the bigger potential factors right away. So Intel's neural compute stick doesn't replace the traditional style of the neural network though. What it does is fill a need on a more basic level, essentially close to the digital the gap between massive and expensive networks and more experimental projects. For example, me and my friend had a little go at running a neural network that allows hand-painted art styles to be transferred to photos.

Starting point is 00:04:13 Lastly, the low-power nature of the Intel neural compute stick opens up even more doors. Yes, it conserves energy, but more important than that, it means compact consumer devices can benefit from new types of AI-powered applications. like virtual and mixed reality, drones, robotics and security systems. Basically, it's neural network computing on the go. A tool like this is powerful and it becomes more so when paired with Intel's neural compute SDK. So, if you'd like to know more about Intel AI, I've included some links in the description below. Anyway, that's all for this video. If you know someone who'd be interested in this, feel free to pass it along to them. Leave a

Starting point is 00:04:53 comment in the description and tell me what you think about neural networks. This has been to GoGo and you've been watching Cold Fusion. I'll see you again soon for the next video. Cheers guys. Have a good one

ColdFusion - Neural Network Prototyping On the Go

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.