Editor's Note: This article was first published on the Text IQ blog.
Children are born with intelligence but lack knowledge. As they observe and interact with the world, experience feeds into that intelligence and, gradually, knowledge takes form. This, of course, is what we call learning.
When we step back to look at it, we see the vast scope and myriad complexities of this simple process that we take for granted. A human must learn an incredible amount of knowledge before ever setting foot in a school, but imagine if we had to teach our children every single one of those things. As it happens, for each concept we teach to them, they learn far more through curiosity, play, and observation—suggesting why unsupervised learning is so essential.
A majority of our knowledge is derived through independent means. Endeavoring to teach either a child or machine enough to conduct higher-order functions like reasoning, planning, and prediction isn’t just inefficient—it’s impossible. For AI to ever close the gap with human intelligence, unsupervised learning is key.
So let’s demystify unsupervised machine learning: how it works, the current state of technology, and why it plays a crucial role in shaping the future of AI.
What is Unsupervised Learning?
Perhaps the easiest way to understand unsupervised learning is in comparison to its anthesis: supervised learning. These systems rely on massive labeled datasets, and generating those labels is a laborious task that requires many human teachers. For instance, training a supervised learning algorithm to spot cancerous tumors necessitates inputting numerous photos of those tumors, and someone must first go through all of those pictures and draw digital circles around the tumors.
So, given enough informative data and a suitable algorithm, supervised learning can effectively solve classification problems and estimate the relationships between variables. Basically, the computer follows a predefined function: If given X input, it yields Y output.
Unsupervised learning, on the other hand, feeds unlabeled data to an algorithm. The machine has no pre-programmed information about the values or output of that data, meaning that we cannot train it to recognize or analyze specific points. Rather, we set the algorithm loose with the opportunity to learn about the data’s underlying structure. As Graves and Clancy from DeepMind put it, unsupervised learning agents “learn for the sake of learning.”
While unsupervised learning may be more challenging than its supervised counterpart, and it’s sometimes difficult to determine if the results are meaningful without access to answer labels, this approach is still necessary despite these issues. Annotating large datasets is costly and time consuming, especially when it comes to massively complex tasks like language recognition. Deciding how to divide the data into appropriate categories can also be daunting.
Today’s Unsupervised Learning Technology
Despite recent innovations in the artificial intelligence space, AI is still in its infancy. While engineers have developed high-performance algorithms for winning Go and scanning documents for identifying personal information, we’re still far from approximating the kind of intelligence that enables a baby to learn about the world and its machinations.
That said, there has been a steady increase in unsupervised learning’s capabilities, and it’s certainly a useful approach. Its primary use is known as clustering, which involves discovering groups of similar examples among unlabeled data to identify structure and determine organization within those collections based on shared similarities.
Clustering helps marketers identify groups of customers with similar behavior; insurance companies to identify fraud; and, of course, companies to maintain privacy compliance by detecting personally identifiable information within huge data sets.
Unsupervised learning is also useful for anomaly detection, or finding unusual data points within a data set, and association mining, or identifying sets of items that frequently appear together. This analysis, in turn, fuels powerful insights that can inform development of better strategies.
One notable technology that utilizes unsupervised learning is the Generative Adversarial Network (GAN), which Yann LeCun, director of AI research at Facebook and professor of computer science at NYU, calls the best idea in artificial intelligence. GAN involves training two neural networks against one another: one generating content and the other attempting to discriminate between the generated data points and those from an actual dataset.
By training together, they become stronger and more accurate. In fact, GAN is the unsupervised learning technology behind the infamous deepfakes, computer-generated faces that look strikingly realistic. Such novel approaches to unsupervised learning continue to expand the realm of possibilities, opening new doors and offering an intriguing glimpse of what is to come.
The Quest for General Intelligence
While enterprises already take advantage of these use cases, researchers see unsupervised learning as a crucial step in the overall AI mission. The pinnacle of artificial intelligence would be to develop a machine with common sense. Although it is possible to design programs with superhuman capabilities for narrowly defined tasks, they’re unequipped to also learn seemingly basic human skills like tying a shoe, catching a ball, reading a book, et cetera.
While a child possesses intuition about common sense, a machine does not. Yet, scientists are striving to create a program that can observe how the world works, interact with it in a meaningful way, and then translate that massive amount of information into predictions, plans, and reasoning. The desired end-game of these efforts is artificial general intelligence, or AGI.
If we can ever hope to program a computer capable of learning without bounds, it must rely on a sophisticated unsupervised learning system. Just as it would never be possible to directly teach a child every concept that they learn and internalize, we cannot hope to do so with AI either.
With all its potential, unsupervised machine learning is one of the most exciting innovations in AI today. On top of that, it’s only one component of the independent learning paradigms that we call “not-supervised learning.” Alongside federated learning and self-supervised machines, unsupervised learning embodies our commitment to building machines that don’t need human labeling to learn how to think for themselves. To learn more about how Text IQ harnesses this technology to detect sensitive information, please reach out to us.
Artwork for this article was created by Sarah Vachlon.