NXP Semiconductors N.V. has rolled out a machine learning environment aimed at developers of intelligent vision, voice, and industrial applications. The new software environment is said to be unique in that it brings machine learning to so-called “edge” applications, where cost, processing power, and memory availability may be constrained. It also simplifies the process of employing machine learning—especially for engineers who have no idea how to get started.
“Up to now, machine learning has been done by major OEMs and big companies on expensive chips,” Gowri Chindalore, head of technology and business strategy for embedded processors at NXP, told Design News. “We are lowering the barrier for implementing machine learning, so it can go into cost-sensitive, edge-processing applications."
NXP’s introduction of the new software environment may be well timed. Machine learning, which is a subset of artificial intelligence that involves pattern recognition, has seen remarkable growth in recent years. International Data Corp. has predicted that spending on AI and machine learning will grow from $12 billion in 2017 to $57 billion in 2021. Similarly, Deloitte Global has said the number of machine learning implementations will double this year (over 2017) and double again by 2020.
At last week’s NXP Connects conference in Santa Clara, CA, engineers demonstrated a microwave oven endowed with machine learning capabilities. The oven was able to recognize foods—hot dogs, waffles, carrots, and broccoli—in about 100 msec using a $3 processor chip. (Image source: Design News)
NXP wants to play a part in that growth by simplifying the implementation of machine learning. It aims to make that implementation possible, not only in $50 high-end application processors, but in $1 microcontrollers as well.
The new software environment could help that happen in two ways, NXP said. The first way is through the application of a software tool that enables developers to quickly assess how well machine learning would perform on their selected processor chip. The tool tells engineers how much time it would take for their processor to make an intelligent decision and how accurate that decision would be.
At Last week’s NXP Connects conference in Santa Clara, CA, the company demonstrated a microwave oven endowed with machine learning capabilities. The oven was able to recognize foods—hot dogs, waffles, carrots, and broccoli—in about 100 msec using a $3 processor chip.
NXP engineers argued that many such applications aren’t using machine learning today because developers believe that machine learning is reserved for higher-end chips. But NXP wants to change that. “Ninety-nine percent of the people who use a microwave take more than a second to close the door,” Chindalore told us. “They don’t need a $15 processor to do that food recognition application in one millisecond, because it makes no difference to the end user.”
The second way in which NXP hopes to help developers is through the application of a smaller inference engine, which can run on lower-end microcontrollers with less on-board memory. “We take the neural net and squeeze the inference engine out, so it’s now right-sized for running on edge devices,” Chindalore said.
Up to now, he said, the burden of extracting the inference engine has fallen on the application developer, who typically has neither the time nor the mathematical expertise to make it happen. That’s why machine learning has typically been relegated to higher-end chips, he added. “What usually comes out is a very large, memory-heavy neural net. As such, it cannot be put into a resource-constrained edge device.”
NXP envisions the new software environment being used in a wide variety of machine learning applications. It could be applied in smart doorbells, for recognizing people at the front door, or in department stories for assessing the emotions of customers. It could also be employed in handheld devices for so-called “wake word” detection or voice recognition. NXP engineers also foresee its use in industrial machinery.
“You could look at the vibration of a cutting tool and warn the owner of the machinery if the machine is vibrating in a weird manner,” Chindalore said.
Ultimately, the company wants to get the word out to the developer community that machine learning isn’t just for expensive, high-end applications. “Up to now, there’s been a chasm between the early adopters and the rest of the developer population,” Chindalore said. “Our goal is to bridge that chasm.”
Senior technical editor Chuck Murray has been writing about technology for 34 years. He joined Design News in 1987, and has covered electronics, automation, fluid power, and auto.