Recent moves by companies like IBM and Google to open-source their machine learning algorithms shine the spotlight on the rising prominence of one of the trendiest areas in computer science today. But what exactly is machine learning anyway? How does it differ from the more familiar concept of artificial intelligence (AI)? And can we ever build truly ‘intelligent’ machines capable of thinking and learning by themselves?

Machine learning’s rise to prominence is a direct result of the explosion in Big Data derived from social media, video, websites, mobile apps and the mass of connected devices that make up the Internet of Things. This explosion of data has forced data scientists to try to find more efficient ways to put that information to use, and machine learning is one way of doing that.

What the hell is machine learning anyway?

In a nutshell, machine learning refers to the study of computer algorithms to provide computer programs with the ability to learn, discover, predict and improve by themselves, just by scanning huge amounts of data, and without any explicit programming, said Dave Schubmehl, research director for Cognitive Systems and Content Analytics at International Data Corp. (IDC). In its simplest sense, it can be thought of as ‘computer programs that learn’.

The process by which computers “learn” is actually pretty simple, Schubmehl said. It starts by feeding the algorithm with sets of training data that’s used as a base to create statistical models. These models are then improved over time with what’s called a “feedback loop” that adjusts models by running them through additional sets of training data. With each successive run, the algorithm improves its performance, first by ingesting the training data and later with production data itself. As a consequence, each successive training and production run introduces new data to the algorithm, thereby helping the program to continually learn, the analyst said.

“At their core, machines learn by generalizing from each experience, applying their generalized hypotheses to new conditions, measuring the performance of each new application of the generalized mathematically expressed hypotheses and modifying their hypotheses before starting another feedback loop,” Schubmehl said.

Data is therefore the lifeblood of machine learning. Without data to “feed” on, machines can’t learn anything new. Conversely, the more data there is, the more “intelligent” the machine can become.

“Machine learning starts with data—the more you have, the better your results are likely to be,” wrote David Chappell of David Chappell Associates, in a recent white paper [PDF] on Microsoft’s Azure machine learning solution.

Surprisingly, the concept has actually been around since before World War II. The first-ever machine learning algorithm is generally accepted to have been the Fisher Linear Discriminant (circa 1936), which Wikipedia describes as “a method used in statistics, pattern recognition and machine learning to find a linear combination of features that characterizes or separates two or more classes of objects or events.”

But although the idea has been around for decades, it’s only recently that enterprises have started to put it to use. That’s because machine learning algorithms demand massive amounts of data and compute power, the scale of which simply wasn’t available at an affordable cost until recently.

“Because we live in the Big Data era, machine learning has become much more popular in the last few years,” Chappell wrote. “Having lots of data to work with in many different areas lets the techniques of machine learning be applied to a much broader set of problems.”

Machine learning vs artificial intelligence

Machine learning is often confused with the much more ambiguous concept of artificial intelligence (AI), and in some cases the terms are mistakenly used interchangeably. But AI covers a much wider spectrum that embodies mimicking human thought, and in truth it doesn’t actually exist. No one has ever developed what could be called true artificial intelligence. Instead, AI is better thought of as the “science” powering its own development.

Sakthi Dasan Sekar, AKA “shakthydoss”, a prominent blogger, programmer, statistician and data scientist who previously worked at Systems Technology Group, Inc., put it this way:

Artificial Intelligence is a science to develop a system or software to mimic how humans respond and behave in a circumference. As a field with extremely broad scope, AI has defined its goal into multiple chunks. Later each chunk has become a separate field of study to solve its problem.

AI encompasses a number of major sub-components, such as natural language processing and computer vision. Machine learning is one of them. But even then, the definition is debatable. In a discussion on the differences between machine learning and AI on Quora.com, Monica Anderson, Founder of Syntience Inc., put forward the argument that machine learning is the only true form of artificial intelligence, even though she concedes there’s a long way to go before machines can really think for themselves.

“As long as the programmer is the one supplying all the intelligence to the system by programming it in as a world model, the system is not really an artificial intelligence. It’s just a program,” Anderson wrote. “The only hope to create intelligent systems is to have the system itself create and maintain its own world models continuously in response to sensory input. Following this line of reasoning, machine learning is NOT a subset of AI. It really is the ONLY kind of AI there is.”

Anderson said that all kinds of machine learning today require some guidance from humans. Building machines that can think without that supervision is “thousands of times more difficult” to do, she wrote.

Building true “intelligence”

If Anderson is correct that even the most promising form of machine learning can’t deliver true AI, it’s logical that machine learning itself will one day be surpassed by even more advanced concepts. One promising technique is “machine intelligence,” which is a concept modeled after the neocortex that mimics the brain’s capability to detect complex patterns.

Machine intelligence is distinguished from machine learning by virtue of being general, self-learning and having on-line training, said Donna Dubinsky, CEO of Numenta Inc. Machine intelligence is “general” because its algorithms don’t change for different solutions, it’s “self-learning” because there’s no supervised training set and it has “on-line learning” because it learns every time it’s fed new data. While machine learning requires building targeted solutions to problems, machine intelligence can be thought of as a broader, more general capability that can be applied to multiple problems at once.

Dubinsky explained that the biggest difference between the two concepts is that machine learning doesn’t really “learn” anything at all, but instead uses its training to recognize patterns and solve problems. “The set of ML algorithms used for any one application are often highly optimized by trained data scientists, and will need to be modified if the application changes,” Dubinsky added.

In contrast, machine intelligence doesn’t require any special “training set” to get started, Dubinsky said. Machine intelligence systems begin by identifying the data patterns in streaming data by themselves, without any tuning or optimization. Recognition of patterns and the ability to predict them follow, and these predictions become more accurate as the system consumes more data.

“If the data pattern changes, [the system] will first recognize this change as an anomaly, and if the data remains in the ‘new’ pattern, it will learn the new ‘normal,’ all without any tuning or optimization by data scientists,” Dubinsky said.

Machine intelligence systems use the same algorithm no matter what task they’re applied to, which means they don’t need to be rebuilt for each new application. Instead, data is encoded into “sparse distributed representations” (SDRs) that are fed into the algorithm, Dubinksy said. This means that only the encoder needs to change when dealing with different types of data, rather than the algorithm itself. That makes machine intelligence quite distinct from today’s machine learning algorithms.

As impressive as it sounds, machine intelligence is still a nascent technology. Numenta has spent more than nine years building its Hierarchical Temporal Memory (HTM) machine intelligence system, but only released its first product last year.

Building machines that are truly capable of thinking for themselves is long and difficult slog, if it’s possible at all. But Numenta co-founder Jeff Hawkins said he believes it’ll be well worth the wait.

“We have to understand our brains if we’re ever going to progress, explore the universe and figure out all the mysteries of science,” he told VentureBeat last year. “I think there can be tremendous societal benefit in machines that learn, as much societal benefit as computers have had over the last 70 years. I feel a sense of historical obligation, almost. How could I not do it?”

Photo Credits: Dick Thomas Johnson via Compfight cc; Blickpixel via pixabay.com; Geralt via pixabay.com