Exploring deep learning, neural networks, and computer vision

(upbeat music) - What's simple and second nature for you can be extremely difficult for even the most advanced computer systems. Neural networks are about taking what you do easily and scaling it up to take advantage of the speed of modern computers. That's why some organizations use neural networks to do a form of unsupervised learning. With unsupervised machine learning, you're letting the neural network cluster your data into different groups. At first glance these clusters might not seem that important, they may even seem trivial, but keep in mind that some of the largest companies are built around creating these clusters. Companies like LinkedIn, Google, Facebook, and Amazon are all using neural networks to cluster together your friends, search history, and buying habits. That's because these neural networks see patterns that would be nearly impossible for any human to find. - The current place where you see neural networks being used the most is in this idea of deep learning, is that you can create a neural network that interacts with lots and lots of data, lots of training data, and it has a number of other properties that get much more technical, like backpropagation and other things. And that allows you to really say, oh, here's how it, it actually has what we typically refer to as lift, benefit in the algorithm. - [Jonathan] Many applications need the ability to process and analyze not only batch data, but also streams of new data in real time. Running on top of Spark, Spark Streaming enables powerful interactive and analytical applications across both streaming and historical data. The real bonus is the fact that you can use virtually the same code that you created for batch data to process real-time data. - [Adam] TensorFlow is a popular tool for building and training deep neural networks. It's used by many companies and research institutions every day to building cutting-edge systems. However, TensorFlow is a low-level tool. It's designed to give you total control over the design of your neural network, but it makes you do a lot of the work on your own pages of code. Keras, on the other hand, is a high-level solution. It's primary design goal is fast and easy experimentation. The idea is that if you spend less time coding, you can spend more time experimenting. Face alignment is where we adjust each face image so that key facial features, like the eyes, nose, and mouth, line up with a predefined template. Here's an example of face alignment. On the left is a face that was extracted directly from a real photo. Notice how the left side of the head is turned away from the camera and how the face is slightly tilted down. On the right is the aligned version of the face. We've adjusted the image so that the eyes and bottom of the nose are as close to the center as we can get them without causing too much distortion. So why do we need to align our face images? We want our face recognition system to work even if the person isn't looking directly into the camera.