hand raised in the middle of a wheat field
PHOTO: Jonny Caspari

As someone working in machine learning, the hype around AI is both exciting and troubling. Although it’s great people want to discuss developments in the field, the way they talk often has the unintended effect of obscuring the meaning of key terms. Whether it’s through the overuse of buzzwords, interchangeably using distinct concepts, or even mistaking simple ML models for a conscious machine, the media spotlight can confuse the topic so much that any discussion becomes meaningless. 

To make matters worse, this happens even as the meaning of those terms are themselves up for debate by experts in the field.

This is anathema to a healthy workflow. My team has found having a defined taxonomy has helped productivity and improved our understanding of machine learning, which has a trickle-down effect on the quality of our work. 

The best place to start building a taxonomy like this is at the very top. Using chatbots as an example, let’s have a look at some of the field’s core terms: artificial intelligence (AI), machine learning and natural language processing (NLP). We’ll explore the subtle differences between them and what this means for your business.

What Is AI?

While many of the finer points are still being thrashed out, the consensus is AI is the oldest and broadest of these three terms. Essentially, the field of AI is concerned with building machines that can perform tasks in a human-like way, and perform these tasks as well as, or better than, most humans. The focus generally concerns tasks where machines have traditionally been inferior to humans, such as communication, reasoning or movement around a space.

The difference between AI algorithms and regular algorithms lies in the way their rules are created. In a normal algorithm, developers set specific rules that define which precise kind of output the software should produce, given a certain kind of input. With machine learning and most AI applications, the software is instead fed many examples of correct input-output combinations and allowed to create its own rules.

It’s very frustrating for researchers when media and the general public conflate AI with the conscious machines they see in science-fiction novels and movies. It is indeed a goal of AI to build machines that are holistically smarter than humans and able to learn new things independently of any human guidance. This is usually referred to as strong AI. However, at the moment it’s unclear whether this is attainable and under what timeframe. 

All AI applications today are actually what is referred to as weak AI. This means the AI application is only learning what it has been told to learn and operating within a very specific field. For example, a chatbot can probably learn new words and sentence structures from language input, but it can’t drive a car or predict the price of shares. As a result, talk of the imminent rise of machines is actually pretty unhelpful. While strong AI or even machine consciousness are interesting topics of discussion, at this point they’re largely irrelevant to much of the actual work being done in the field.

Having said that, AI research posits that if your description of the task is precise enough, machines can be built to simulate any feature of intelligence. It’s therefore unsurprising that attempts to improve the ability of machines in human-dominated areas are usually focused on building and developing algorithms that tackle specialized tasks. These algorithms use a range of fast processing techniques to understand patterns and relationships behind annotated datasets. If the task is well-defined and the data quality is high enough, this often results in algorithms that can outstrip human performance.

With all this in mind, it’s obvious that chatbots and virtual assistants can be classified as AI. Their ability to understand speech, act upon it, and generate a response closely resembles the way that humans communicate with each other. Most chatbots are able to have this impact because of the machine-learning processes they’re built upon — let’s explore that next.

Related Article: 5 Drivers of Personalized Experiences: A Walk Through the AI Food Chain

What Is Machine Learning?

Although many people use the two terms interchangeably, machine learning is generally considered to be a subset of AI. It’s the predominant way we build AI applications. While ‘artificial intelligence’ describes the goal, ‘machine learning’ describes the models, processes and supporting technology that we’ve been using in our attempts to create intelligent machines. The field’s close relationship with data mining and analytics means it can also be considered a form of applied statistics.

Although AI leaves the door open for other paths to machine intelligence, most advances towards this goal so far have been made using machine-learning algorithms. These have some key characteristics that separate them from other algorithms, and that will define the field if another route to AI is discovered in the near future. Machine learning is primarily concerned with algorithms that can make connections between various annotated data and their output. Crucially, they are also able to learn independently from new, varied output, thereby improving their models without the need for human intervention. This approach lends itself to many of AI’s defining use cases, such as computer vision and machine translation.

It’s debatable whether any AI applications to date haven’t derived from machine learning in some way. Almost all current chatbots have been built by machine learning, but there is another approach that some data scientists are considering. Rule-based models are founded on linguistic systems that are developed by experts to imitate the ways humans structure their speech. When a system is supported by a comprehensive set of human-built rules like this, it can be difficult to say whether the system is truly machine learning. However, diving into the debate around this particular issue might make the differences between AI and machine learning a little clearer.

As we’ve seen, the line between these two concepts can be blurry. It’s probably easiest to consider AI as the goal and machine learning as the algorithms and processes that are getting us there, while remaining open to the idea that this may change in the future. Further to this, it’s also possible to break down AI along thematic lines. One example of this is natural language processing.

Related Article: How to Differentiate Machine Learning From Dressed-Up BI

What Is Natural Language Processing?

This field covers efforts to replicate one particular area of human-like AI tasks in machines. This is predominantly done via machine learning but can theoretically be achieved through different methods. NLP covers most processes that involve communication by machines. This includes any attempt by a machine to recognize what is being said, understand its meaning, determine the correct course of action, and respond to the user.

Chatbots are a useful example of an NLP application that spans the entire communication process. Through the example of one interaction with a chatbot, many of the main challenges of NLP come into view:

  • By making a request from a chatbot, either through text or through speech, you engage Natural Language Understanding (NLU) algorithms. This is concerned with understanding the intended meaning of what is said. It does this through turning raw text into structured data that it can make sense of.
  • Using this data, it creates commands for itself and processes the data in a decision engine.
  • Through Natural Language Generation (NLG), a chatbot will turn its results from data back into text that you can understand. The main difficulty here is in creating speech that sounds natural, while also clearly communicating the information the user needs.

The vast majority of these processes involve a lot of machine learning, since understanding language requires algorithms that can understand a vast and wildly varied amount of input. However, it’s entirely possible that machine learning is not essential to effective NLP. Regardless of this, all elements of NLP can also be classified as AI. Whether the focus is on entity extraction or text generation, the main goal is to improve a machine’s ability to communicate — and therefore perform a human-like task. As long as this is the case, it falls under AI’s vast and vague umbrella.

Related Article: Calling All Linguists: The Messaging Bots Need Help

What Does This Mean For Your Business?

Whether you work in marketing, sales or project management, the hype around AI will inevitably reach your office. When it does, you and your management team will have to decide how to react. With a clear understanding of the relationships between these key AI concepts, you’ll be able to ignore some of the more outlandish pipe dreams and instead focus on the potential ways that AI can help your business.