Machine learning is hot—and it’s moving on from pattern recognition and traditional algorithms into more elusive “deep learning.” What does it mean for IT pros to operate in a world where their computers don’t always need to be pointed in a particular direction, they only need data to figure it out on their own? We’ll take a look in a few blog posts.
What is Machine Learning?
Machine learning draws on this idea: computers can learn from data alone through an iterative process by which models are exposed to new data sets and evolve independent adaptations without reprogramming.
With the expansion of computing power — not to mention the exponential increase in available data and affordable, scalable storage — it’s now possible to automatically produce models to analyze ever larger quantities of information and deliver increasingly accurate predictive results in faster and faster cycles. This is a welcome development for businesses, because with more precise modeling comes the ability to identify profitable opportunities and risks before they happen.
Outside the IT field, “learning computers” and the artificial intelligence they are fueling sound like the stuff of science fiction, but IT pros know machine learning is already touching lives in many ways. There are self-driving cars, both the full version in testing and the various features included in current production models, such as collision avoidance, smooth cruise control, and self-parking. There are the online recommendation engines behind Netflix and Amazon, which use machine learning to serve more relevant content and product choices. And there are banking and credit card fraud detection programs, Siri’s automated task management capabilities, and the list goes one.
Machine learning is here, and it’s only going to expand.
Types of Machine Learning
Not all machine learning is the same, however. There are many approaches, which have different applications. For example:
- Supervised learning involves data that has been categorized. As ExtremeTech explains, if data showing apartment rents and floor space was fed into a supervised learning scenario, it would graph the relationship between the two factors. The resulting algorithm could later predict the rent of any apartment based on the floor space, “without writing a specific program to perform the same task.” The process learns and adjusts as it returns “correct” and “incorrect” results based on the data provided, increasing accuracy and self-updating as the relationship changes. The method is useful to process historical data to develop future predictions.
- Unsupervised learning uses data where no “right answer” is known. The goal is more exploratory, to find as yet undiscovered relationships or structure within the data. A common example is using retail transactions to find audiences that could be treated similarly for marketing purposes. Or in the rent example, unsupervised learning could look into a wide range of factors—local property values, crime, demographics, school quality, etc.—and find cluster relationships between these factors and rent prices.
- Semisupervised learning uses labeled and unlabeled data. It can save money by using a small amount of expensive, labeled data to “train” the system and apply the results to the more inexpensive, unlabeled data.
- Finally, reinforcement learning is a trial-and-error process in which the learner aims to maximize rewards, resulting in a policy for optimizing results.
What’s New Here?
Some aspects of machine learning may sound familiar from when data mining was the rage. In fact, data mining is similar in its intent to glean insights from raw information, but not all of the methods used in data mining are machine learning. The field also draws on statistics, text analysis, time series analysis, and other means. A key contribution of machine learning is its ability to “discover” structure within data using an iterative approach, without needing humans to begin with any theories or assumptions to test.
This is a big part of why deep learning has become the latest buzzword, because it takes machine learning’s greatest asset to the extreme. Relying on neural networks, deep learning identifies more complicated patterns in larger amounts of data than previously possible. It holds great promise for language translation, medical diagnosis, and much more.
Will that bring us to true AI? Maybe, but that’s not the only goal. As one commentator explains “Artificial Intelligence is the broader concept of machines being able to carry out tasks in a way that we would consider ‘smart,’ and Machine Learning is a current application of AI based around the idea that we should really just be able to give machines access to data and let them learn for themselves.”
Deep learning can certainly be applied to the group of challenges we view as AI, the ability to mimic and interact in a natural way with humans. But machine learning and deep learning will also serve many other purposes in the future, which may or may not involve AI. For the latter, think of the “Jack of all Trades,” IBM’s Watson, deriving insights from huge amounts of data and then communicating it in a way humans readily understand.
Or if you want to stay up at night, think of the movie Ex Machina. Maybe Elon Musk has watched it a few too many times.
Chris Adams is President and COO of Park Place Technologies. Contact him at firstname.lastname@example.org.