your123 -

How to build scalable edge AI systems

A look at the different challenges enterprises and vendors face in this new arena, and some of the different ways the merging technology can be applied, including in healthcare.

More and more enterprises are looking into ways they can apply edge AI to their systems.

The new spectrum of AI is quickly attracting the interest of vendors and enterprises users.

The meaning of edge

Edge itself is a term that's still being defined.

Speaking during a panel session of the Edge AI Summit on Nov. 15, Brandon Gilles, CEO of spatial AI vendor Luxonis Holding Corporation, defined edge as computing locations that don't fall under data centers.

Building on that definition, Moshe Mishali, co-founder and CEO of Deep AI, said whatever is not on the cloud or in a data center is edge. Edge can include desktops, laptops, small servers, sensors and  applications that don't run on a central hub.

Mishali added that many of Deep AI's customers use what they call an edge data center.

As a vendor that's focused on edge and hybrid data systems, Hewlett Packard Enterprise defines edge as where new data meets a decision-making model, said Sreenivas Rangan Sukumar, distinguished technologist at HPE.

"Edge means computing under constraints," Sukumar said.

That might mean making decisions when a user doesn't have enough time to process it offline, when connectivity is intermittent or asynchronous, or when other limitations are present.

Another, more obvious definition of edge AI is the combination of edge computing and AI, Mishali said.

Image of a panelist discussion at Edge AI summit 2021
Panelists at Edge AI summit discuss what Edge AI is, the challenges and different applications.

Challenges in scaling edge

One challenge in scaling edge is that it's nearly everywhere, compared to the more centralized cloud, according to Mishali.

Another challenge is training edge AI models.

"The approach today for edge AI is to do the inference very close to the data source," Mishali said.

Edge means computing under constraints.
Rangan SukumarDistinguished technologist, Hewlett Packard Enterprise

He added that for training, data needs to be captured and then sent to a central data center or to the cloud. This requires a lot of power, and a lot of GPUs, to do effective training.

"This can be a real challenge for scaling applications," Mishali continued, adding that enterprises face the difficult challenge of making sure that data flows smoothly among numerous endpoints and bandwidths.

Deploying AI on different edges

Yet another challenge is the diversity of different edges. The question for vendors and enterprise users is how they can deploy AI on the different edges.

For enterprises looking to use AI on different edges, Sukumar said there are questions they must ask themselves, including how complex the decision-making model is, whether it must be deployed on the edge, and whether the organization wants to track or measure the accuracy of decisions.

"So given my data input sizes, given my model input sizes, given my power constraints, given my bandwidth constraints, is the system going to do the job that I expected it to do?" Sukumar said.

His advice for enterprises looking to build models on the edge is to "start small, and then build towards scale, rather than thinking toward scale first."

Edge AI applications

While a familiar application of edge AI is autonomous vehicles, it's not the only one.

A key application is in the mobile area.

"That's where edge AI was most beneficial the fastest," Gilles said. That's because much of what new smartphones are doing -- from speech recognition to personalization -- has a touch of AI. Healthcare is also an important field for edge AI, to monitor the health of newborn babies, for example.

"With the combination of edge AI, spatial sensing, and computer vision, you can have the equivalent observational power of a trained doctor to see how a baby is moving," he said.

Edge AI systems can alert healthcare providers if a newborn baby exhibits certain changes that indicate a neurological development disorder, such as lack of motion.

"So, with the edge you can now deploy this outside the NICU [neonatal intensive care unit]," Gilles continued. "So, if you have a high-risk baby, you can have better than NICU-level monitoring as if the world's best doctor was staring at the data everywhere, because you're quantizing that data and then you can get it down to a text message size."

Other applications of edge AI are in the manufacturing, retail and physical security industries.

"It's things that humans can't do without this," he said. "When you put all these edge systems out, then humans can do a lot more amazing things because they have the data that they would have to consume more than all of their time to produce if they were doing it manually."

Next Steps

Empower intelligence at the edge with AI deployment

Dig Deeper on AI technologies

Business Analytics
Data Management