Andrew Ng uses a rubber sheet analogy to explain how data augmentation works: collecting more data for one type of input (like cafe noise in speech recognition) lifts performance in that region and nearby areas. Iterative error analysis identifies the next biggest gap to target, making this an efficient path toward human-level performance.
There's a picture, a conceptual picture that I found useful for thinking about data augmentation and how this can help the performance of a learning algorithm. Let me share this picture of you, since I think you find it useful too, when trying to decide whether to use data augmentation. Take speech recognition. There could be many different types of noise in speech input, such as car noise, plane noise, train noise, machine noise, or cafe noise, or library noise, which isn't that loud, or food court noise. Maybe these types of noises are more similar to each other because they're all mechanical types of noise. And these types of noise may be a little bit more similar to each other with mainly people talking and interacting with each other. So let me share with you a picture that I keep in mind when I'm planning out my activities on getting more data through data augmentation or through actual data collection of any of these types of data. In this diagram, the vertical axis represents performance, say, accuracy. And on the horizontal axis, and this is a conceptual kind of a thought experiment type of axis, I'm going to represent the space of possible inputs. So, for example, there's speech with car noise and play noise and train noise sound a bit like car noise. So they're quite similar. And machine noise is a little bit further away by machine noise. I'm picturing the sounds of a washing machine or a very loud air conditioner, say. Then you may have speech with cafe noise, library noise, or food court noise. And those are maybe more similar to each other than to these types of mechanical noise. Your system will have different levels of performance on these, different types of input. Let's say the performance is this for data of plane noise, that of car noise, train noise, machine noise, and it does worse on data with liptunity for improvement. Now, what happens if you use data augmentation, or maybe not data augmentation, but go out to a bunch of actual cafes to collect a lot more data with cafe noise in the background. What you would do is you would take this point. Imagine grabbing a hold of this blue rubber band or this rubber sheet and pulling it upward like so. That's what you're doing. If you collect or somehow get more data with cafe noise and add that to your training set, you're pulling up the performance of the algorithm on inputs with caffeine noise. And what that will tend to do is pull up this rubber sheet in the adjacent region as well. So if performance on caffeine noise goes up, probably performance on the nearby points will go up too. And performance on faraway points may or may not go up as much. It turns out that for unstructured data problems, pulling up one piece of this rubber sheet is unlikely to cause a different piece of the rubber sheet to dip down really far below. Instead, pulling up one point causes nearby points to be pulled up quite a lot. And far away points may be pulled up a little bit, or if you're lucky, maybe more than a little bit. But when I'm planning how to improve my learning algorithms performance and where I hope to get it to, and getting more data in those places to iteratively pull up with those pieces or those parts of the rubber sheet to get them closer to human level performance. And when you pull up part of the rubber sheet, the location of the biggest gap may shift to somewhere else and error analysis will tell you what is the location of this new biggest gap. That may then be worth your effort to collect more data on and therefore to try to pull up one piece at a time. And this turns out to be a pretty efficient way to decide where on the blue rubber sheet to pull up next to try to get performance closer to, say, human level performance. I hope this analogy of a rubber band or a rubber sheet and repeatedly pulling up a point on this rubber sheet will help you predict the effects of collecting more data that's associated with a specific category or a specific tag. How do you get more of this data? Let's take a look at how you can perform data augmentation and some best practices of doing so in the next video.
