Behind the Scenes: AI

In the previous post, I expressed my distrust of theories claiming that AI robots will one day destroy humanity. This article will be more scientific-based as I dive into how artificially intelligent machines operate.

At the origins of AI study, two approaches stood out: symbolic AI and sub-symbolic AI. In symbolic AI, scientists use rules, words, and phrases to teach a machine to perform an action given a scenario. It’s like a conditional if/then statement that you write in experimental labs. Let’s use self-driving cars as an example. If a self-driving car sees a stop sign, then it will stop. Now imagine programming rules for every possible scenario that a car could face. A deer could cross a road, a deer and kid could simultaneously cross a road, a branch could fall from a tree, etc. The possibilities are endless! So while symbolic AI machines can be used in limited domains to execute a set number of functions, they would never be able to perform all the tasks that humans do on a daily basis. Instead, computer scientists have reached the consensus that subsymbolic AI machines have the best chance of reaching human intelligence. Researchers in this camp derive their approaches from neuroscience. Specifically, they believe that AI networks should be modeled on our brain’s neural networks. This has led to the idea of Deep Learning: the training of deep neural networks. The most common type of deep neural network is called a Convolutional Neural Network (ConvNet for short). This may get a little confusing so hang on!

A ConvNet consists of multiple layers of neural networks that each play a distinct role in performing a specific task. For example, let’s say that we had to use a ConvNet to recognize images of an animal. First, the image’s details — numbers associated with individual pixels — are inputted into the network. The first layer of the ConvNet evaluates a specific characteristic of the animal such as its edges. The next layer will then measure a different feature such as the animal’s legs and wings. And so on the process goes with each subsequent layer responsible for identifying an even more specific detail. Finally, once the image’s properties have been assessed by many layers (could be 100+), the image reaches the last layer which is called the “classification module” (Mitchell 73-74). Here the ConvNet will give a percentage probability that the image is a certain animal. For instance, if the ConvNet has been trained to recognize pictures of eagles and falcons, the classification module may conclude a 70% chance that the image is an eagle and a 30% chance that it is a falcon. Now you’re probably wondering how the network can make these predictions. How does a ConvNet know what an eagle should look like? Frankly, it’s a very simple but tedious process. Initially, a scientist can program the network to recognize certain basic features of an eagle or a falcon. However, there is no chance that humans could account for every characteristic especially when images are evaluated based on the content of their pixels (for reference, a 4*6 inch image can have upward of 200,000 pixels). Instead of manually programming the network, the network learns by itself.

Or that’s what the media would want you to believe. It’s actually a combination of human assistance and a process called back-propagation. I’ll start with human labor. Let’s go back to the eagle and falcon scenario. To help the ConvNet discern between the two animals, scientists feed the network hundreds of data sets that have pictures of both eagles and falcons. As mentioned in the previous paragraph, the ConvNet will assign percentages for each image. But then, in a manner very similar to Operant Conditioning in Psychology, scientists will tell the machine whether their guess was correct or not. The researchers’ job is done. Back-propagation kicks in. Back-propagation is a programmed algorithm that the ConvNet uses to fix its own errors (Mitchell 38). When told “wrong” for an image, the ConvNet will work backwards — hence the title back-propagation — and adjust each layer to account for the mistake. The actual math behind this technique is very complex and so I won’t delve into that. All you need to know is that through the back-propagation algorithm, a ConvNet can fix its different layers to produce the correct result. When back-propagation occurs many times over a large set of pictures, the ConvNet moves closer to pinpoint accuracy. Of course no AI machine is 100% accurate in their task, but there have been stark improvements with the introduction of ConvNets and back-propagation; success that validates the continual use of ConvNets in future AI projects.

In this post, I used image-recognition as an example of the elaborate efforts undertaken to produce a successful AI technology. I said it’s a tedious process because of the amount of time scientists invest in gathering the thousands of pictures that help refine the ConvNets. But remember this machine can only do one task: recognizing images. So you can imagine that it would take an infinite amount of resources to construct an AI computer capable of doing an array of objectives at human-level.

The last thing I will leave you with is a recommendation to check out Melanie Mitchell’s novel Artificial Intelligence: A Guide for Thinking Humans. All of the information presented in the previous article and this article have come from Mitchell’s book. She goes into much more depth than I do, but manages to explain complex ideas with user-friendly jargon. Mitchell explores several other applications of AI that you may find interesting: speech-recognition — used in Alexa, Siri, and Google Translate — and the ubiquitous topic, self-driving cars.

Citations

Mitchell, Melanie. Artificial Intelligence: a Guide for Thinking Humans. Farrar, Straus, and Giroux, 2019.