The Machines are Learning, and Humans Should Understand How
What is machine learning, and how does it work? And what are AI, deep learning, and neural networks? Here’s a primer with everything you need to know.
There’s no doubt that the age of AI has arrived, and it’s here to stay. A cursory glance at the headlines on any given day are enough to support this idea; they are constantly filled with news of self-driving cars, AI-powered machine translation, learning cybersecurity systems, and more. Google, Facebook, and almost every other Silicon Valley giant are pouring funds into AI research.
Amidst all of this fervor, however, very few are stopping to discuss what artificial intelligence actually is. This makes it difficult to join the conversations surrounding AI and understand the excitement behind the technology. So here, we’re going to break down the basics of AI and machine learning and how they work.
To start, let’s take a step back and consider artificial intelligence (AI) as a whole. There’s a great deal of conflicting information about what is and isn’t artificial intelligence. Simply put, artificial intelligence is the idea of creating a machine or system capable of emulating human intelligence.
If that definition sounds vague or even somewhat subjective, that’s because it is. There is no one decisive interpretation of what constitutes “human-like intelligence,” or how a machine should exhibit it. Fifty years ago, products like Microsoft Excel might have been considered AI for their ability to calculate complex math problems. The target for what we consider to be AI is constantly moving, as the term “AI” can refer to a broad range of ideas and technologies, from ad targeting on Facebook to smart manufacturing equipment.
Machine Learning: The New “AI”
Today, AI is often used as a synonym for machine learning, a subset of AI that has recently seen a rapid increase in technological progress and user interest. Traditionally, computer systems were programmed with a strict set of rules, and could not operate outside of those rules. Those systems are often referred to as If-This-Then-That programs. Machine learning systems, however, are designed to continually learn and adapt beyond their initial programming input. This allows machine learning systems to understand and perform tasks of a complexity far greater than is possible for any traditional system.
Take language, for example—a deeply human enterprise that has, in the past, been notoriously difficult for machines to comprehend. Researchers have long worked to create a computer system that can use and understand language in a human-like fashion. The problem is that human languages are massive and vastly complex systems, the precise rules of which researchers even now are still trying to ascertain. How can we possibly encode a program with every single rule of a system when we don’t even know all of the rules ourselves?
Machine learning is helping us make huge strides in understanding this problem. Humans may not know all the rules of language, but we still learn them all starting in infancy, merely from continued exposure. Machine learning has made huge leaps forward in natural language processing by taking the same tack with computer programs: instead of trying to teach a program every single rule manually, it can be exposed to a sufficiently large corpus of language and eventually learn the rules itself.
This, then, is why machine learning is so often associated with “big data.” Machine learning programs typically require large datasets to learn from in order to be accurate and precise. If a dataset is too small, the model can develop a skewed and inaccurate understanding of the problem.
Does Your Machine Need Supervision?
There are two ways to approach machine learning: supervised learning and unsupervised learning. In supervised learning, the algorithm is initially fed pre-labeled training data. For example, if you wanted to train a supervised model to correctly identify different kinds of fruit, you would first feed it a dataset of fruit that have already been identified. By seeing the fruit paired with the correct labels, the model learns what categories of fruit exist and how to recognize them. It can then use this information to identify unlabeled fruit in the future.
In unsupervised learning, on the other hand, there is no training data. You would simply feed the model an unlabeled dataset of fruit and let it determine on its own how to categorize the fruit, be it by shape, color, size, or any other characteristics. Rather than labelling a given fruit an “apple,” the model would then cluster together fruit with similar characteristics that it has determined most likely belong in the same group. As one might guess, unsupervised learning models are trickier to extract value from, but they are capable of handling large and more complex problems with little information up front.
Not all machine learning systems are strictly supervised or unsupervised—rather than a binary, this is often a scale, where systems may lean more heavily on one approach or the other but make use of both in some way or another.
Deep Learning and Neural Networks
This brings us to another set of terms that are thrown about liberally but rarely defined: neural networks and deep learning.
Neural networks are a type of machine learning that string multiple functions together. The type of neural network most commonly applied at present (and certainly the type most commonly discussed) is deep learning. Deep learning also strings multiple functions together, but it does so in hierarchical layers, allowing for a greater level of complexity in its analyses.
To illustrate how this works, we can return to our fruit classifier. The first layer of a neural network intended to classify fruit would be a fairly simple model that would learn to recognize very simple features, such as the edge of a shape. The next layer would then use the output from the previous layer to recognize more complicated features, such as how those edges fit together into corners. In this way, the neural network can use successive layers of increasing complexity to approach even very large problems in a highly nuanced and detailed fashion.
The development of techniques such as deep learning have allowed for a new generation of AI; one that can function at levels of competency and even intelligence that could only have been dreamed of even a decade ago. These are the computer systems that are driving cars and defeating human champions in games like Go. Humans have dreamed for decades of creating intelligent machines, but it was not until the advent of machine learning and its associated techniques that research could not only leap forward, but do so in the continuous and sustained manner we see now. It’s critical to start learning the basics of AI and machine learning because these technologies are reshaping every human enterprise, and they’re not slowing down anytime soon.
At SparkCognition, we employ machine learning for cybersecurity, predictive maintenance, natural language processing, and more. For more information, visit https://www.sparkcognition.com/ or contact us at firstname.lastname@example.org.