Machine learning is one of the hottest trends in technology today. In fact, Gartner put machine learning at the peak of its most recent Hype Cycle for Emerging Technology. And the firm has predicted that by 2020, artificial intelligence (AI) technologies, including machine learning “will be virtually pervasive in almost every new software product and service.”
According to IDC, organizations will spend $12.5 billion on AI systems in 2017. That’s a huge 59.3 percent increase over 2016 levels, and the analysts say that spending will continue to grow at more than 50 percent per year through 2020. At that point, total AI revenue could top $46 million. David Schubmehl, research director, cognitive systems and content analytics at IDC, said, “Cognitive/AI systems are quickly becoming a key part of IT infrastructure and all enterprises need to understand and plan for the adoption and use of these technologies in their organizations.”
So what is machine learning? What is its relationship to artificial intelligence? And what should technology professionals know about its potential benefits and challenges?
What Is Machine Learning?
The first person ever to use the phrase “machine learning” was likely Arthur Samuel, who developed one of the first computer programs for playing checkers. In 1959, he defined machine learning as technology that gives “computers the ability to learn without being explicitly programmed.” Other computer scientists have proposed more mathematical definitions for machine learning, but Samuel’s definition remains one of the most accurate and easiest to understand.
Machine learning is a subset of artificial intelligence, the segment of computer science that focuses on creating computers that think the way that humans think. In other words, all machine learning systems are AI systems, but not all AI systems have machine learning capabilities.
You can subdivide machine learning into several different categories:
- Supervised learning requires a programmer or teacher who offers of examples of which inputs line up with which outputs. For example, if you wanted to use supervised learning to teach a computer to recognize pictures of cats, you would provide it with a whole bunch of images, some which were labelled as “cats” and some of which were labelled as “not cats.” The machine learning algorithms would help the system learn to generalize the concepts so that it could identify cats in images it hadn’t encountered before.
- Unsupervised learning requires the system to develop its own conclusions from a given data set. For example, if you had a large set of online sales data, you could use unsupervised learning to find clusters or associations among that data that could help you improve your marketing. You might discover, for instance, that women born in the early 1980s with incomes over $50K have an affinity for particular brand of chocolate bar or that people who buy a certain brand of soda also buy a certain brand of chips.
- Semi-supervised learning, as you probably guessed, is a combination of supervised and unsupervised learning. Going back to the cat example, imagine that you have a large number of images, some of which have been labeled “cat” and “not cat,” and some of which haven’t. A semi-supervised learning system would use the labeled images to make some guesses about which of the unlabeled images include cats. The best guesses would then be fed back into the system to help it improve its capabilities, and the cycle would continue.
- Reinforcement learning involves a system receiving feedback analogous to punishments and rewards. A classic example of reinforcement learning (as it applies to machine learning) is a gambler sitting in front of a row of slot machines. At first, the gambler doesn’t know which slots will pay off or how well, so he tries them all. Over time, he discovers that some of the machines are set “looser,” so that they pay off more frequently and in higher amounts. As time passed, the gambler — or in this case, the computer program — would increase his earnings by playing the looser machines more often.
Machine Learning Use Cases
Organizations in a wide variety of industries have already begun experimenting with machine learning. In some cases, software vendors have incorporated machine learning into tools used for a specific purpose, and in other cases, users have adapted general-purpose machine learning applications for their needs. Some of the most common use cases for the technology include the following:
- Fraud detection — Banks and credit card issuers have been among the first to use machine learning. They often use the technology to identify transactions that might be fraudulent. If your credit card issuer calls you to see if you recently made a particular purchase, the company probably used machine learning to flag a suspicious transaction on your account.
- Recommendation Engines — The online recommendation engines used by companies like Amazon and Netflix are among the most familiar examples of machine learning. Using data gleaned from millions of shoppers and users, the machine learning systems are able to predict items that you might like based on your past shopping or viewing habits.
- Search — Google, Microsoft Bing and other search engines use machine learning to improve their capabilities on a minute-by-minute basis. They can analyze data about which links users click in response to queries in order to improve their results. They are also using machine learning to improve their natural language processing and deliver specific answers to some questions.
- Video surveillance — Machine learning is enabling facial recognition systems to improve all the time. In some cases, these systems can identify known criminals, or it may be able to identify behavior or activities that are outside of the norm or break the law.
- Handwriting recognition — The US Postal Service uses machine learning to train its system that recognizes handwritten addresses.
- Natural language processing — Today, most of us take for granted that personal assistants like Siri, Cortana or Google Assistant will be able to understand voice requests and respond to questions. Over time, these tools use machine learning to improve their abilities to recognize, understand and process verbal input.
- Customer service bots — Automated agents can use natural language processing and customer service data to answer common questions and improve the quality of those answers over time.
- IT security — Many of today’s most cutting-edge IT security solutions, like user and entity behavior analysis (UEBA) tools, use machine learning algorithms to identify potential attacks. In the case of UEBA, machine learning establishes a baseline of “normal” behavior that it uses to detect anomalies, potentially allowing organizations to identify and mitigate zero-day threats.
- Streaming analytics — In today’s 24/7 world, a lot of data, such as social media feeds and online sales transactions, gets updated constantly. Organizations use machine learning to find insights or identify potential problems in real time.
- Predictive maintenance — The Internet of Things (IoT) offers many potential machine learning use cases, including predictive maintenance. Enterprises can use historical equipment data to forecast when machinery is likely to fail, enabling them to make repairs or install replacement parts proactively before it impacts business or factory operations.
- Anomaly detection — In much the same way that machine learning can identify anomalous behavior in IT systems, it can also detect anomalies in manufactured products or food items. Instead of hiring inspectors to examine goods visually, factories can use machine learning systems that have been trained to identify items that fail to meet standards or specifications.
- Demand forecasting — In many industries, getting the right amount of product to the right location is critical for business success. Machine learning systems can use historical data to predict sales far more accurately and quickly than humans can on their own.
- Logistics — For transportation companies, setting up schedules and routes is a complex, time-consuming chore. Machine learning systems can help identify the most efficient and cost-effective way to get goods or people from point A to point B.
- Financial trading — Every trader hopes to find patterns in the market that will allow him or her to buy low and sell high. Machine learning algorithms can help identify potential opportunities based on past market activity.
- Healthcare diagnostics — Many experts envision a future where machine learning diagnostic tools work alongside human professionals to identify diseases and determine the most effective course of treatment. Computer systems may be particularly good at detecting anomalies in various kinds of scans and in spotting rare diseases.
- Self-driving cars — Autonomous vehicles are one of the most fascinating applications of machine learning. In the not-too-distant future, vehicles capable of navigating on their own may become the norm.
- Robots — While they have long been a staple of science fiction, robots with machine learning capabilities could very soon be a part of everyday life. These robots would be able to improve their capabilities over time, allowing them to become more useful to humans.
Source: IDC Worldwide Semiannual Cognitive Artificial Intelligence Systems Spending Guide
Machine Learning Benefits
Many of the use cases described above can be handled by humans or software without machine learning capabilities. However, machine learning technology offers several benefits over each of these alternatives:
- Speed— Humans can create the models, input the data and run the calculations necessary for predictive analytics on their own. However, humans — or humans using software without AI capabilities — might need days, weeks or months to accomplish tasks that machine learning tools can complete in just seconds, minutes or hours.
- Accuracy — That speed allows machine learning systems to utilize a larger volume of data and a larger number of models than humans ever could. As a result, AI systems are much better than people at some tasks, such as predictive analytics. However, in other areas, such as voice recognition or image recognition, computer systems still haven’t achieved the same level of accuracy as human beings.
- Efficiency and cost savings — Machine learning software isn’t cheap; in fact, in some cases it can be very expensive. However, it is often far more affordable to use software to automate a chore than to hire dozens or hundreds of people to complete the same task.
Machine Learning Challenges
While machine learning has a lot of potential and is already becoming commonplace, the field faces a lot of challenges — some organizational, some technological and some philosophical.
- Talent scarcity — Enterprises often need data scientists to operate their machine learning systems, and workers with these skills have become among the most highly sought after. Their salaries are some of the highest in the technology industry, and in recent years, their average pay has been rising very quickly. However, data shows that this trend may be slowing as vendors roll out machine learning software with self-service capabilities that make it possible for non-data scientists to use it.
- Lack of data-driven culture — While most executives understand the potential benefits of data-driven decision-making and machine learning technology, getting everyone in a large enterprise to change their mindset and activities is often a long, slow process. Machine learning advocates frequently face internal hurdles when trying to promote the technology.
- Poor data quality — The best AI systems in the world cannot return good predictions and insights if the data feeding their models is inaccurate. Many organizations find that they need to improve their data cleansing and data management processes before they can fully utilize machine learning software.
- Data integration — At many organizations, data still resides in siloed applications and storage solutions. Feeding all that disparate data into a machine learning system can pose a challenge, but vendors are responding with solutions that can accept a wide variety of data types and formats.
- Data security — Balancing the need to restrict access to data with the need to use data to feed machine learning systems can be tricky. Some organizations may need to update their policies and/or use machine learning tools that encrypt or anonymize data.
- Infrastructure requirements — Advanced machine learning systems run best on hardware with multiple, fast CPUs and GPUs. In addition, they require a lot of storage space and appropriate networking capabilities to move the data from storage to applications and back again.
- Ethical dilemmas — AI is becoming more like human beings, but it lacks the sense of morality that informs most human decision making. For example, when Microsoft released a social media bot named Tay that had machine learning capabilities, it quickly learned to say inappropriate and offensive things. Some experts are calling for technology companies to make sure that AI systems follow a strict set of ethical rules in order to prevent them from committing crimes, harming humans or even wiping out the human race.
- Fear — Many people find the idea of artificial intelligence in general or machine learning in particular unsettling. They worry that computers will take over their jobs — with good reason. Forrester has predicted that “cognitive technologies such as robots, artificial intelligence (AI), machine learning, and automation will replace 7 percent of US jobs by 2025.” Others, including Tesla and SpaceX CEO Elon Musk, worry that machine learning could pose an existential threat to humanity. Whether or not those fears are well-founded, organizations will have to find a way to deal with them if they want to experience the potential benefits of machine learning.