The promise of AI

- [Narrator] When people talk about artificial intelligence, they can get that glassy, far-off look in their eyes, and have visions of machines that take care of the drudgery and the work of life, like watering your garden. But I'd like to deal a little more with the reality of artificial intelligence, and one thing that can help with that is defining a little bit what it is that we're talking about. So for example, artificial intelligence, or AI, is a general term that refers to the idea of thinking machines, or computers that can learn from experience and operate without specific programmed instructions. So some of those tasks can involve visual perception, logical reasoning, and learning in general. But AI is not an especially specific term. so let's compare and contrast with a few other terms that cover some of the same intellectual and technical terrain. Probably, the most important of these is machine learning, or ML, and this refers to computer algorithms that adapt to data and predict values or classes. Now, machine learning algorithms range from the simple to the extraordinarily complex. The simple ones can include a basic linear regression, which sometimes you can even do by hand. And the complex can include neural networks, I'll say more about those in a minute. But it's a very large category, machine learning, and one of the neat things about it is many of these algorithms adapt to prediction errors. So the algorithm runs through, it makes your predictions, and then it sees which cases it misclassified or it was far off, and then it can either adapt the algorithm or create a secondary algorithm for dealing with those cases. In fact, it's possible to create many models or algorithms, and then combine those using what's called an ensemble approach, and that is one of the things that gives machine learning it's extraordinary power. One particular method, or algorithm that is frequently used, is neural networks, or NN, actually should be artificial neural networks, or ANN. But this refers to a complex kind of algorithm that has an input layer where the data comes in, shown here on the left, and it has one or more hidden layers, we've got two hidden layers in this one, and then we have an output layer where you get the prediction, the result, on the far right, and you can see that they are densely interconnected. Now, not only are they densely interconnected like the neurons in the human brain, one of the important things is that they have a nonlinear activation function. So it's not simply additive, there's sort of a triggering that goes on, again, analogous to what happens with the human brain. And what this does is it gives it a lot of flexibility in the data that it can deal with. Also, there are many different kinds of neural networks for different prediction challenges. So for instance, you can use a convolutional neural network for image data, or a recurrent neural network for time series data, or a recursive neural network for hierarchical tree organized data. There are many, many more options than that, but this has been one of the great areas of machine learning and AI. And if you want to be a little more specific, there is also deep learning. And these are neural networks with many hidden layers, so hundreds or thousands or millions of layers. And so, while that may feel like it is simply a quantitative difference between the standard neural networks, it gives something that is qualitatively distinct in its performance. In fact, it's these deep learning algorithms that have led to so many of the newsworthy abilities of AI. On the other hand, because they are so extraordinarily complex, and because of the non-linear connections between things, they can be a little bit like a black box, meaning, you know the data you fed in, and then you know that it did something, and then you got your predictions, which are often very accurate, but it's hard to trace the data all the way through, and this presents certain challenges for policy and legislation that we'll talk about later in the course. Also, deep learning neural networks, especially large, production-level ones, can use a very significant, huge amount of energy to run them. This is a major challenge in the growth of the area, and finding ways to get the benefits without using quite such an extraordinary amount of energy in the process. But as you might guess, because there have been so many amazing developments from deep learning, and because we still have some of these challenges, this is an enormous growth area for research and development. On the other hand, I want to mention there is also the, quote, unquote, bread and butter, predictive modeling and analytics. Pie charts and bar charts are still extraordinarily useful for people who are making data-driven decisions, but it's a different thing from artificial intelligence. This is an area of great value, I strongly encourage people to be familiar with it and to use it when it's appropriate. And then for the situations where something more complex and more nuanced is necessary, artificial intelligence could be a great solution. I also want to mention a few things about the timeline of AI, because it hasn't been consistent, it has gone in fits and starts, where you have periods of development, and then periods that are referred to as AI winters, where very little happened. Modern AI began in the 1950s, where work began to create machines that could reason, reach conclusions, make decisions, and learn from mistakes, just like humans. It was a very exciting, very promising time, but it quickly ran into some major obstacles that led to the work going basically into hibernation in the '60s. But in the '70s, more progress was made by drawing from game theory and from experimental psychology, so taking a whole different paradigm. In the 1990s, something happened that nobody ever thought a computer would be able to do. IBM's computer, Deep Blue, beat the world chess champion Gary Kasparov. Chess is a very complex game, and this was often taken as one of the standards for what it meant to be intelligent and creative, and the fact that a computer was able to do it was a huge achievement, and really a place marker in the development of AI. Then in the 2010s, deep learning became economically feasible due to changes in hardware, so distributed computing, and in the software, the programming that made it possible. And that led to the 2020s, where AI has become widespread, and both the promises and the challenges have become more evident to a wider range of people. Not surprisingly, laws and policies have been developed, and are continuing to develop to make it so that people can harness the power of AI and circumvent some of the problems that can come along with it. And so, we're still not to robots and flying cars and superhuman intelligences, but the AI landscape is developing rapidly, and we have both the promises, and some of the perils, that we will be discussing in this course as a way, again, of getting the benefits of the AI revolution, and avoiding some of the most common pitfalls.