On the intelligent use of AI

Artificial intelligence has dominated the world of technology recently. But what really is Artificial Intelligence and how often is the term misapplied?

Artificial intelligence has been one of the main research topics in computer science for decades. As far back as 1950, Alan Turing devised his famous test (or Imitation Game) designed to tell if a computer was able to show intelligent behavior. Over recent years we have got closer and closer to true artificial intelligence. However, most current forms of artificial intelligence are narrow. That is, they only exhibit intelligence in a very specific field.

However, anyone reading the popular press, or looking at company hype would think we are already in the era of full artificial intelligence. Companies globally are claiming to use AI to solve every sort of problem. Indeed, AI has now become such an overused term that it is starting to become devalued. So, how many of these companies claiming to do AI can really justify the claim? And what is artificial intelligence anyway?

Background

Everyone has heard of artificial intelligence. But what actually does the term mean? Of course, there are plenty of formal dictionary definitions such as this one from Merriam Webster

Artificial intelligence

1: a branch of computer science dealing with the simulation of intelligent behavior in computers

2: the capability of a machine to imitate intelligent human behavior

However, these definitions aren't all that helpful. So, how do AI researchers define things? Well, they generally talk about "Narrow AI" and "General AI". A narrow AI is any computer system that applies some form of intelligence (usually learning) in order to complete a task. A general AI is one that is capable of acting autonomously, applying many forms of intelligence to solve a problem it hasn't seen before. Let's look at these terms in more detail.

Narrow AI

A narrow AI applies intelligence to solve a single, specific problem. The most common forms of narrow AI are all based on machine learning (ML). Here, machines learn to spot patterns in data by one of 3 approaches:

• Supervised learning, where the computer is trained using carefully labeled data
• Unsupervised learning, where the computer learns by finding patterns and creating hypotheses
• Reinforcement learning, where the computer is rewarded for making correct decisions

Other common forms of narrow AI include:

Computer vision. Here, a computer uses approaches such as recurrent neural networks to locate and identify objects within still and moving images. This is essential for things like self-driving cars and face recognition.

Natural language processing (NLP). This is where a computer is taught how to parse natural language and extract the actual meaning. This is used in virtual assistants such as Apple's Siri and Amazon's Alexa.

Deep learning. This specialized form of machine learning has received a lot of attention recently. Here the computer is able to teach itself. One of the most famous examples is Google Deep Mind teaching itself to play Go.

AI often excels at these sorts of tasks. It has been said that AIs like these possess superpowers because they often outperform humans at tasks we feel we should be good at.

Artificial general intelligence

Ask any member of the public to describe what artificial intelligence is and the chances are they will mention something like Skynet. These dystopian visions of all-powerful global super-intelligence have been mainstays of science fiction for years. However, AIs such as these are still very much in the realms of science fiction. At least they are for now. AIs like this are described as exhibiting general intelligence. That is, they apply multiple forms of intelligence to solve a wide set of problems.

Importantly, general AIs don't need to be trained to solve a specific problem. The term for this is artificial general intelligence. For years, we believed this form of intelligence set us apart from animals. However, there is ample evidence that some animals display such intelligence. You can find films of crows learning how to make tools and even rats learning to drive. As things stand, no computer can exhibit this sort of intelligence. However, deep learning is the closest we have got yet.

What do advertisers mean by AI?

It's no secret that AI is a buzzword that sells. Over the past few years, almost every software company has started to use the term in its marketing. Managers use it in project outlines in an effort to secure budgets. Startups use it when pitching to venture capital firms. Software engineers are desperate to find excuses to try it in their code. The buzz around AI has caused a huge shortage of data scientists. And universities have been quick to try and plug the gap, launching huge numbers of specialized courses in data science, machine learning, and related topics.

But how many of the systems that claim to use AI are really using it? And how often is it just another marketing term? Even if the company is using AI, are they using it in an intelligent manner? Is AI really giving better performance than a traditional algorithm would? AI-purists would say that no system can really justify the use of the term yet. But what if you are one of the majority who accepts that narrow AI is still AI?

You might think it's simple. If you use ML, computer vision, NLP, or deep learning you have an AI system. But just dumbly using one of these techniques doesn't necessarily equate to being artificially intelligent. All too often, it feels like companies are adding AI functionality as something of a gimmick. Or the AI doesn't really function that well, and they have to admit to getting humans to help. Most cloud providers offer marketplaces where you can buy AI functions as a service and add them to your product. You could argue that, while the underlying image recognition, ML or NLP are forms of AI, a piece of software that just uses that service is probably not a form of AI. At best, I'd argue that such services should describe themselves as AI-powered.

What should AI mean when a company uses the term?

If AI-powered services and products aren't actually AIs themselves, when is it OK to use the term? I believe there are 2 important aspects to consider. Firstly, the product should be significantly enhanced by using AI. That is, the AI should actually perform a critical role, either improving efficiency or providing better performance. Often, managers that turn to AI as a magic solution for business intelligence are disappointed because the models don't perform as well as expected. Secondly, the service shouldn't just be repackaging an underlying AI application from a cloud marketplace. Or if it is, it should be adapting it for the specific use case being solved.

So, what do I mean when I talk about intelligent AI? Here is a simple checklist for anyone looking to invest in an AI product or project:

• Make sure the problem is actually suitable for AI. Many projects fail because of a lack of data, meaning you can never create a strong model.
• Check that the data you collect is suitable for AI techniques. As an example, medical observations are often hard to apply ML to because they happen at varying time scales.
• Don't just repackage an existing model unless you are certain how it will perform in all cases. In other words, test the model extensively with your own data.
• Test the results of the AI system against the existing approach and make sure it gives better outcomes. Sometimes you are better off using classical statistical analysis and data science.
• Look at combining multiple forms of AI. Read up about the concept of "boosting" ML models. Even consider hybrid approaches where you combine data science with AI.

If you can answer these questions in the positive then you are probably justified in using AI. In turn, your marketing team can feel reasonably confident about using the term too. And then your only problem is responding to the people that think AI is just a meaningless buzzword!

##

About the Author

Jon Seaton is the Director of Data Science at Functionize. Jon has worked with many Fortune 500 companies on machine learning and artificial intelligent initiatives and his background spans a range of industries from medical to finance. He was a triple major in mathematics, physics, and engineering at UW-Madison and worked at the MIT Media Lab in graduate school. Jon has a passion for entrepreneurship and is actively involved in volunteering his time and energy towards helping startups succeed.