What Machine Learning Needs To Learn Next

Machine Learning – Today, consumers can buy cameras powered by artificial intelligence (AI) that recognize a person at their front door. But they don’t have anything close to a robot that can tie their shoelaces. Why? Simply put, because most machine learning algorithms available today in AI applications don’t learn very well.

Thanks to a branch of AI called unsupervised learning, however, that’s about to change.

Most machine learning uses a technique called supervised learning. It recognizes objects (such as a picture of a cat) by looking at lots of examples (cat pictures). It also looks at lots of pictures of things that aren’t cats to help it distinguish.

People must find and label thousands of pictures to train supervised learning algorithms. “The humans are the slow and expensive part of this process right now,” says Amir Zamir, a postdoctoral researcher at Stanford University and UC Berkeley.

When the supervised AI finishes learning, researchers wind up with something that can recognize a domestic cat. But if you want to recognize a dog or even a close relative like a tiger, you have to start all over again.

“For every new task you need to learn, you often have to relabel your data,” says Derik Pridmore, CEO of Osaro, which uses AI to help robots pick and place items in warehouses.

The Limits Of Supervision

Another problem with supervised learning is that it makes embarrassing mistakes. In one instance, an AI traffic cam publicly shamed a Chinese billionaire after spotting her jaywalking in the middle of a road. Except she wasn’t—the traffic cam saw her face pictured on the side of a bus. The Chinese AI recognized a face but didn’t understand the implications of an image on the side of a bus.

“When you analyze those mistakes, you realize that mostly it’s because the models don’t understand the world,” says Yoshua Bengio, a professor in the Department of Computer Science and Operations Research at Université de Montréal.

What if we could learn without labels? That’s the promise of AI that goes beyond supervised learning. There are many branches of it, each solving different problems, but one thing they have in common is less reliance on labeling, says Zamir.

“The ultimate case is where there’s no human at all, and there’s nothing annotated. You just receive a bunch of raw images,” he says. Then, the AI starts noticing things itself.

For example, researchers might feed thousands of images of house interiors into their training models. A class of unsupervised learning known as clustering may notice a common shape in many of the pictures, explains Zamir.

“The fact that they are repetitive means the AI can cluster them together,” he says. The AI has identified TVs, even though it doesn’t know what they are.

This article originally appeared on forbes.com To read the full article, click here.

Nastel Technologies uses machine learning to detect anomalies, behavior and sentiment, accelerate decisions, satisfy customers, innovate continuously.  To answer business-centric questions and provide actionable guidance for decision-makers, Nastel’s AutoPilot® for Analytics fuses:

  • Advanced predictive anomaly detection, Bayesian Classification and other machine learning algorithms
  • Raw information handling and analytics speed
  • End-to-end business transaction tracking that spans technologies, tiers, and organizations
  • Intuitive, easy-to-use data visualizations and dashboards

If you would like to learn more, click here.