Writing new rules: How machine learning enables computers to teach themselves
Browse Blog Topics

Writing new rules: How machine learning enables computers to teach themselves

Intelligent machines are making inroads into a wide array of applications but before they really become pervasive, they must achieve a critical capability: the ability to learn.

One of the most important techniques advancing AI is actually something quite simple: learning from examples. Machine learning has evolved to the point where humans are no longer just feeding detailed instructions to machines, but training them with massive amounts of data.

“Today’s AI is really machine learning,” says David Eun, president of Samsung NEXT and Chief Innovation Officer of Samsung Electronics. “This is an approach where you gather lots and lots of data, and you crush it together, and you come up with pretty compelling outputs.”

Those compelling outcomes range from creating empathetic robots trained to improve elder care to using data from the Internet of Things (IoT) to optimize municipal services in smart cities. The End of the Beginning video series from Samsung NEXT explores how machines are learning to teach themselves.

Machine learning is more than programming
There is an important distinction between machine learning and programming. Programming is a set of detailed instructions that humans give machines so that they can carry out specific instructions.

Machine learning, on the other hand, occurs when machines analyze and discover patterns in data that allow them to adapt to varying circumstances without any changes in instruction from humans.

“It’s no longer engineers writing code, for example, to detect objects in a video. It’s the data that is writing the code,” says Adi Pinhas, CEO of Brodmann17, a company specializing in computer vision.

Today’s ML algorithms are tackling problems that have vexed machine learning researchers for decades, such as translating languages with human-level accuracy. Consider the range of problems that are solved with ML: Some algorithms learn how to classify things, from fraudulent credit card transactions to cancerous tumors. Other types of algorithms are designed to spot groups of related things, such as customer segments to discovering genes with similar functions.

A trio of machine learning approaches
Most successful applications of machine learning are based on only three types of algorithms: supervised, unsupervised, and reinforcement learning. One aspect of artificial intelligence is being able to distinguish different types of things.

“Let’s say I want to build a machine that identifies cats. I give it an image, and it needs to tell me is there a cat in that image,” says Yosi Taguri, founder of MissingLink.ai, which is a product of Samsung NEXT. “How do you write the rules of how you identify a cat? You take a lot of images with cats, a lot of images without cats, and every time you feed it an image with a cat it learns something, and every time you feed it an image without a cat it learns another thing. If you’re doing it enough times, the machine suddenly understands what makes a cat a cat.”

Unsupervised learning operates with information about the data, for example, if there is a cat in an image or not. This form of machine learning is useful for finding subgroups or clusters within data, such as customer segments, or detecting anomalies like a fraudulent credit card transaction.

Reinforcement learning uses interactions with the environment to develop models. Intuition Robotics companion robot ElliQ, for example, learns by interacting with humans and assessing their responses to the robot’s actions.

“We’re going to see the basic generation one voice assistant evolve to something else which is much smarter, uses context, adapts to our personality, and it uses layers of emotional intelligence, in addition to just basic artificial intelligence, to drive the behavior models behind it.” says Dor Skuler, CEO of Intuition Robotics.

Bringing machine intelligence to smart cities
Machine learning is a core technology facilitating IoT and driving the development of smart cities. Sensors distributed across a city generate far more data than could be efficiently analyzed humans but machine learning algorithms can extract valuable information and help municipalities optimize the use of civic resources. “Having these smart cameras in the city that can stream data all the time … about what is happening can enable other systems to be able to prioritize and to tune the city.” according to Ahi Pinhas.

The potential for growth and diversification in machine learning is quickly becoming more apparent for smart cities and beyond. As David Eun predicts, “We’re going to see applications of AI in fields that we didn’t think possible, fields where human creativity was really driving it, but the kind of outputs will be startling. And we’re beginning to see that now.”

For more information on what’s possible with machine learning and where it’s going next, look into the video series by Samsung NEXT, The End of the Beginning.

Related Stories