It’s tempting to think that the artificial intelligence revolution is coming — for good or ill — and that AI will soon be baked into every facet of our lives. With generative AI tools suddenly available to anyone and seemingly every company scrambling to leverage AI for their business, it can feel like the AI-dominated future is just over the horizon.

The truth is, that future is already here. Most of us just didn’t notice.

Every time you unlock your smartphone or computer with a face scan or fingerprint. Every time your car alerts you that you’re straying from your lane or automatically adjusts your cruise control speed. Every time you ask Siri for directions or Alexa to turn on some music. Every time you start typing in the Google search box and suggestions or the outright answer to your question appear. Every time Netflix recommends what you should watch next. 

All driven by AI. And all a regular part of most people’s days.

But what is “artificial intelligence”? What about “machine learning” and “algorithms”? How are they different and how do they work?

We asked two of the many Georgia Tech engineers working in these areas to help us understand the basic concepts so we’re all better prepared for the AI future — er, present.

Romberg is the Schlumberger Professor in the School of Electrical and Computer Engineering (ECE) and senior associate director of Georgia Tech’s Center for Machine Learning. He noted there’s no set definition of the term “artificial intelligence” and most researchers treat this idea on a case-by-case basis when they’re deciding if something should be considered AI.

Romberg said the AI decision-making process, at its core, is just like any other calculation that a computer makes. Some combination of data is fed into the system, there are constraints that the algorithm or device operates under, and a result is produced.

And this is where engineering and science bend a bit toward philosophy.

“What you can’t escape is that, ultimately, everything we call an AI is really just a very concrete computational algorithm that takes in some input and spits out some output,” Romberg said. “The real question is, is that what our brains do?”

AI vs. Machine Learning

For the non-experts among us, these terms can be conflated, sometimes used together or interchangeably. They are different concepts, however.

“Machine learning refers to a family of techniques or an entire discipline on how you learn from data,” Romberg said. “Obviously, learning from data is a big part of artificial intelligence, but there are other things you might call artificial intelligence that don’t learn.”

For example, machines that play chess might be considered AI, but these systems aren’t really learning from data. Rather, they’re very good at rapidly exploring a whole range of possible scenarios in the game to identify the next move that most likely leads to victory.

“A lot of the higher-level applications of achieving different forms of artificial intelligence are built on top of machine learning,” said Xie, who works specifically in this area. “Machine learning is more like the foundation, but developing machine learning algorithms involves many other foundational scientific pieces, including mathematics, statistics, combinatorics, optimization, and many more.”

The first step to creating an AI application is defining what the tool needs to achieve. Once developers understand that, they can collect data — lots of data — and create a model. The model is usually an abstract way of representing the data, like a statistical model or sequence modeling. (As an illustration, sequence models can be used for language. Sentences are ultimately a sequence of words, with dependencies and grammar that help shape meaning.) Researchers use a variety of common modeling approaches for this work.

“Once we have a handle on that model, a lot of times after that is math — algorithms and implementation coded up into some pipeline,” Xie said. “Then we listen to the machine to actually have something.”

Training the Algorithms

Training algorithms to take data and produce results is painstaking work. But at its simplest, Romberg said, it’s a math problem: fitting a function to data. 

He used the example of image recognition and creating an AI tool to sort images of cars, trucks, and airplanes. The algorithm takes in a photo, proceeds through a series of computations where different weights are applied to the image’s pixel data, and the data is combined in a variety of ways, and then it produces a single value: car, truck, or airplane.

Training the system to correctly identify the cars as cars, and not as airplanes, requires teaching it with many examples.

“What you try to do based on all of those examples is adjust the weights so you’re getting the right answer for each of the examples you give it,” Romberg said.

In a neural network, algorithms are layered atop one another to process data and pass on the important parts to a higher level. The approach is taken from one model of how brains function. In very simple terms, some stimulus activates neurons, which then feed data to each other and combine it in different ways. Once the information reaches a certain threshold, the information passes to the next layer of processing and so on.

Neural networks work similarly, collecting, weighing, and passing along data in a hierarchy from bottom to top.

“The lower level feeds forward to the next node, and then you combine the data, passing through an activation function. And the combination also has weights attached to it. These are going to select which information is most important,” Xie said. “When you design these algorithms, you have parameters that are the weights and activation and many, many layers. This is such a flexible architecture.”

Interestingly, both Xie and Romberg noted it’s not always clear why a neural network or other kinds of AI algorithms actually work. The complexity of the layers and the millions or even billions of parameters involved can make it challenging to understand why an algorithm or a neural network of algorithms produces a result — even if it’s the correct result. This is an area both Xie and Romberg are working to untangle in various ways.

“One interesting thing about AI and machine learning is that it’s been a highly experimental science so far: people have tried techniques out and they work. Some of that has refuted, or bumped up against, how we understand classical statistics,” Romberg said. “Some of the work I do is trying to understand how AI algorithms really work. Can we put them into a classical mathematical framework?”

Xie likewise uses statistics, machine learning, and optimization principles to shed light on the functions of tools like neural networks so scientists can build better ones — and trust the output of such systems.

Why Now?

Even in the days when computers far less powerful than our smartphones took up entire rooms, scientists started to wonder, “What if?”

“Since the 1950s, when computers were first invented, people have been thinking about how to do artificial intelligence,” Xie said. “It’s not a new thing.”

So why does it seem like AI is now suddenly everywhere? Probably thanks to a confluence of factors, not least of which are the internet and social media. When OpenAI released its ChatGPT chatbot in November 2022, it quickly went viral. Now companies are rushing to bake AI functions into existing tools and develop new ones.

“ChatGPT basically showed many people something they hadn’t seen before about how to interact with a machine,” Romberg said. “It’s a very tangible illustration or tangible packaging of research, and specifically research in large language models.”

The current awareness of AI tools — and specifically generative AI for chats or images — is less a watershed moment in research or AI development, though. It’s more about how relatable those tools were to a broad audience.

“The ability to answer queries is just part of a continuing trend of things we’ve seen. We’ve had recommender systems for Amazon and Netflix. We have Google that prioritizes our searches. We’ve had Siri that gives us at least semi-coherent answers to the questions we ask,” Romberg said. “So, this has been building over time.”

The visuals in this feature were created using Adobe Firefly’s text-to-image generative AI tool.