AI for beginners: What are artificial intelligence, machine learning, deep learning, and NLP?

    Here’s a challenge: define artificial intelligence and list some examples of AI technology. Then ask your coworker, a friend, or a stranger to do the same.

    I’ll bet your definitions and examples of AI don’t match. They may be closely related, but it’s pretty typical for people to have different takes on what does and doesn’t count as AI.

    Trying to define artificial intelligence

    As more AI-based and AI-laced technology is developed every year — now to the point where it’s everywhere you look — people argue over what “counts” as artificial intelligence.

    The simplest definitions say that AI is engineering that enables machines to exhibit human intelligence. (This leads to arguments over what counts as intelligence, but no need to go there now.)

    “Artificial intelligence is a branch of computer science dealing with the simulation of intelligent behavior in computers.”

    To get a bit more specific, AI is often described as goal-oriented technology. It processes, learns, and applies knowledge in order to accomplish a specific task, such as improving writing. AI with machine learning (which I’ll explain more in a moment) can become more precise and accurate as it completes a task repeatedly — just like a human.

    “Artificial intelligence is a system’s ability to correctly interpret external data, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation.”
    Andreas Kaplan and Michael Haenlein

    The late Professor John McCarthy, considered one of the founders of the discipline, said that AI is bigger than us. It can be taught to simulate human knowledge and then expand into tasks beyond our capabilities. For example: quickly processing and finding patterns across billions of pieces of information.

    “Artificial intelligence is related to the similar task of using computers to understand human intelligence, but AI does not have to confine itself to methods that are biologically observable.”
    Professor John McCarthy

    To me, the explosion of the term “AI” is a marketing phenomenon. And what is considered to be AI changes over time (more on that later). Artificial intelligence is more complex than a single definition. In fact, it’s a huge field that doesn’t mean much until you break it down much further. 

    So rather than trying to define AI, it’s actually more helpful and accurate to think of artificial intelligence as “The Goal”. That means that, for one reason or another, practitioners of AI want to construct intelligent machines. They try to meet different subgoals (sometimes called subdisciplines) and a variety of different techniques to achieve The Goal. 

    Which is why you’ll also hear many other terms and acronyms tossed around when people discuss AI. So let’s dive into a few subgoals and techniques: machine learning, deep learning, and natural language processing.

    ML, DL, NLP, and other fun acronyms

    There are several distinct, evolving ideas within the AI discipline. I’m going to focus on the relationships between artificial intelligence, machine learning, deep learning, and natural language processing because they are most relevant to understanding Writer’s work in AI.

    Here’s a picture of the AI family:

    AI family by Writer
    The AI family. (Illustration by Writer, the most advanced AI writing assistant.)

    There are several things to learn from this picture:

    What is artificial intelligence?

    We’ve already gone through the definitions above. What’s most important to note from this picture is that artificial intelligence is a goal. The big, all-encompassing box. It’s what we are trying to accomplish through all this work.

    What is machine learning?

    Machine learning, or ML, is a technique used to achieve AI goals. 

    We give the computer data and a way to know that it’s on the right track. This is frequently in the form of labels that say “this input should produce this output”. For an AI Writing Assistant like Writer, this data might look like “this sentence has a mistake in it, the corrected version looks like this”. However, for a game-playing AI, you might tell it, “see this number up here in the corner? That’s your ‘score’ – make that number go up.”

    It’s like saying, “Mx. Computer, here is training data (i.e., a map of inputs to their correct outputs). Please figure out the ‘program’ being used to build those mappings (i.e., the mathematical function that will match inputs to their correct outputs).” Once the computer figures out that program, it can apply it again to new data to solve new problems. 

    Machine learning is currently dominating all areas of AI.

    What is deep learning?

    Deep learning, or DL, is an approach to machine learning, or a way of teaching the machines. 

    With traditional approaches to ML, programmers spend a lot of time improving models using very fancy and complex theory. Even after all that work, they would often hit a wall at a certain point, and be unable to do better at a goal. Traditional approaches to ML include work like conditional random fields, or CRF and support vector machines, or SVM

    Deep learning is cool because it seems to be infinitely scalable. If you want to improve the model, you make the model bigger (add more neurons) and feed it more data. This is, of course, a simplification; there are nuances to the connections between neurons, and every few months there’s a revolution in our understanding about this. However, the fact that it is even approximately true is remarkable.

    Breakthroughs in deep learning are driving the current boom in AI, and the majority — I’d guess 95% — of research in AI is currently happening here. (The other 5% are contrarians doing contrarian things.)

    Here’s an amusing picture that illustrates traditional machine learning approaches versus deep learning. In the top panel, you see what a manager of a team of people doing traditional ML would say (or, as far as I can tell). Those are all real things that improve models. But none of those models do as well as deep learning, even though there’s far less theory behind deep learning. Deep learning, even without fancy theories, still blows everything else out of the water. (Predictably, in the real world, the two camps don’t agree with the others’ methods.)

    Traditional machine learning v. deep learning. (A meme from Calculated Content)

    What is natural language processing?

    Natural language processing (NLP) is a subset of AI goals

    The goal is to enable computers to process human language, primarily composed of natural language understanding, or NLU, and natural language generation, or NLG. Most people and organizations — Writer included for our AI writing assistant — are doing research in ML to meet their NLP goals. 

    To give you examples of other subsets of AI goals, I’ve included a few in the diagram: advanced audio/visual processing and generation, autonomous robots/cars, computer vision. Just like NLP, they all lean heavily on machine learning and deep learning.

    AI and the future of writing

    So, what is Writer’s goal? To build the best AI writing assistant, enabling everyone in your company to write better, using a consistent style, terminology, and voice. Our goal has various sub-goals: 

    • Identifying and correcting spelling and grammar mistakes
    • Distinguishing between different writing styles and voices
    • Finding ways to change the style of writing to be more consistent / on-brand without changing the meaning
    • And more!

    Now that AI helps people write better and save time in the process, AI writing assistants have evolved beyond basic spell-check and grammar-check. AI now helps people quickly identify poor quality writing or writing that may be misconstrued. AI quickly scans copy to flag all the possible errors, in tone, meaning, punctuation, and more. 

    And not just students and writers use AI for writing and editing. Now that just about everyone is a writer, more and more top brands rely on sophisticated AI like Writer. Our writing assistant helps ensure that everyone who writes anything on behalf of their company uses a consistent voice, aligned terminology, and follows their messaging guidelines. Brands are taking advantage of Writer’s capability to incorporate business brand guidelines and review everything your company writes, from website copy to emails, and promotional materials.

    Writer’s AI also now checks for plain language, inclusivity, and plagiarism. We designed Writer to help businesses scale their communications processes, and we’re just getting started.

    Putting it all together

    Your major takeaway here: know that AI (whatever that happens to mean this decade) is The Goal. Currently, machine learning techniques are your best bet for achieving any AI goal — more specifically, the deep learning type of ML.

    If you want to dig a little deeper, here’s a good Q&A with Ilya Sutskever, the chief scientist at OpenAI, from 2018. Listen from at least 51:20 to 53:29 to dig into more of what I was talking about above about the infinite scalability of Deep Learning.

    A moment of silence for the victims of AI Effect

    Now that you understand more about AI, maybe you’re rethinking what you consider to be “true AI technology”. 

    My advice: if you want to draw a line between “This is AI” and “That is just a computer being a computer”, be sure to put a date on it. Because though your claim may be true today, public opinion will probably change in five years.

    Consider spell checkers. The first spell checkers were simple (by today’s standards) rule-based programs — if this, then that. For instance, if a typed word does not exist in the reference dictionary, then underline the word (assumed typo) in red. Not particularly intelligent when compared to a self-driving car and therefore not AI, right?

    The answer to that question entirely depends on the year you’re living in. Before spell check systems existed, only educated people could identify writing mistakes, and fixing spelling errors was entirely dependent on human intelligence. 

    In the 1960s, the first English language spell checker program was released, and suddenly computers could also identify errors and provide corrections. If you were alive in 1960, you certainly considered this technology to have artificial intelligence. And, wouldn’t you know, it came to us from Stanford University’s Artificial Intelligence Laboratory.

    Today, spell checkers are more complex; they can also process context and give us context-based spelling corrections (e.g. knowing when to use “you’re” vs “your”). But the reason we no longer associate spell check with AI is because it has become a victim of AI effect. This is a naturally occurring phenomenon in which technology that was once considered AI no longer is — essentially due to the lack of novelty and sheer boredom with an idea. 

    These days it’s more shocking not to have spell check than it is to have it. Since spell check  has become “too solved” a problem, it’s no longer considered AI. Thanks to the AI effect, I half-jokingly, half-seriously define AI as “whatever computers can’t do reliably yet.”

    “Every time we figure out a piece of [AI], it stops being magical; we say, ‘Oh, that’s just computation.”
    Dr. Rodney Brooks, researcher and roboticist 

    Spell check isn’t alone. There are plenty of examples of technology that have fallen from grace due to AI Effect: 

    • Web search–related matches (“That’s just Google being Google!”)
    • Document retrieval (“Of course I should be able to search my computer to find my document!”)
    • Software-controlled opponents in games (computerized chess was the pinnacle of AI at one point)
    • Automated song and movie recommendations

    These were all once just scientists’ dreams, and now they’re commonplace technology. The point is that once AI effect takes over, the layman’s understanding of AI changes — hence, the definition is a moving target.

    Finding common ground to define artificial intelligence

    All this to say, artificial intelligence is a nebulous space and the discipline evolves dramatically every year. What we consider advanced applications of AI today could be just a computer doing normal computer things by next year.

    Researching new applications of AI are fascinating and fun challenges to tackle. The Writer data science team is excited to be bringing our AI writing assistant technology to improve content writing and strategy. If you have any questions about what we’re working on, Tweet at us @Writer.

    Sam Havens is Writer’s Director of Data Science. He leads the team that builds the AI in Writer’s AI writing assistant.