Artificial intelligence is typically trained on servers, and it may not be able to learn when running on a less-powerful edge device with limited network connectivity. Last week, Neurala Inc. announced that its Brain Builder AI vision platform has been optimized for edge learning, which it said can be useful for robots and other devices in manufacturing and visual inspection.
Deep neural networks (DNNs) are often unable to recognize new or varied items coming out of a production line, especially as product cycles accelerate, according to Neurala. The Boston-based company said the Brain Builder software development kit (SDK) allows DNNs to be quickly modified to recognize a new product at the compute edge without having to go back to a server.
“Traditional approaches to training DNNs often fall short in deployment when the network encounters a new situation at the edge that it was not trained to classify,” stated Massimiliano Versace, co-founder and CEO of Neurala. “That’s why Neurala has been developing our Brain Builder SDK, which enables users to continue training and tweaking a DNN even after initial training.”
The latest Brain Builder SDK debuted as a partner of Bosch ConnectedExperience (BCX), Europe’s largest Internet of Things (IoT) hackathon, which took place at Bosch ConnectedWorld (BCW) in Berlin. More than 700 developers used Bosh IoT Suite services and tools including the Brain Builder SDK to create prototypes of IoT systems. They worked with devices including cameras and sensors in cars, robots, and more.
Lifelong learning for deep neural networks
“Neurala was funded in 2006, and some of its earliest work was on edge projects for various government research institutes,” said Daniel Glasser, vice president of customer success at Neurala. “In the three years I’ve been here, the most frequent request is, ‘Where’s my data? How do I make sure it stays private and safe?'”
“There’s a great need for edge computing on smart devices, phones, or in manufacturing and automation. Data processing needs to happen locally,” he told Robotics Business Review. “That’s why we’re focusing on edge AI. A lot of people are looking only at the analysis, but Neurala can do training at the edge as well.”
“With Lifelong-Deep Neural Network, or LDNN, you can train AI systems with less data. Instead of training on 50,000 images, you could use a few hundred, depending on the system,” Glasser said. “Then processing requirements drop, and you may not need a server farm. AI can be trained in a fraction of the time on a smartphone or a GPU on the manufacturing floor.”
How does Neurala’s Brain Builder compare with server-based AI? “We’ve taken similar small data sets, and we typically outperform them,” replied Glasser. “With big data sets, performance will be close. We’re ultimately very competitive with DNNs.”
Brain Builder at the edge
“The initial release of the Brain Builder platform was in March of last year,” Glasser said. “Non-experts can train end-to-end vision systems through a Web portal, and everything happened in the cloud. With the upgraded SDK, everything that was done in the cloud can now be done at the edge.”
“Neurala partnered with Bosch to integrate into an affiliate’s safety ecosystem,” he said. “It used Brain Builder to build a brain and deploy at the edge on security cameras.”
“Beyond that, we’ve built systems for drone operations company AviSight to run our AI during electrical infrastructure inspections,” added Glasser. “The drone can point out defects or broken components in real time, without an Internet connection. The processor isn’t on the drone, but it is in a field unit.”
IoT security and mobility with Brain Builder
While 5G networks promise greater bandwidth and lower latency, the usefulness of edge processing for industrial IoT will not diminish anytime soon, Glasser said.
“There are questions abut how reliable it will be, and even the best bandwidth in the world doesn’t answer questions about privacy and the cloud,” he said. “Autonomous vehicles, drones, and delivery robots have high safety requirements. They don’t want to worry about connectivity, so they’ll need edge AI.”
Adding flexibility to machine vision, robots
Neurala has focused on autonomous mobile robots (AMRs) rather than self-driving vehicles because it wanted use cases that can be deployed quickly, explained Glasser. However, the company has “had some conversations” around vehicles, he said.
“Brain Builder is applicable to mobile robots in warehouses,” he said. “Not just for edge computing, but they can also learn incrementally. This is different from DNNs, which are limited to what you first trained them on. Now, a camera for quality inspection or on an AMR can learn something new about a product or a piece of equipment. It doesn’t have to start from scratch.”
“With Neurala’s technology, you can show it a new thing, give it a name, and then teach that robot how it should respond,” Glasser said. “This saves time and cost and adds flexibility.”
“For object recognition, we did a project with a major warehouse logistics provider that had 2 million SKUs to pick and place,” he said. “You can train most AI to recognize tens of thousands of things, but then it tries to fit everything into those boxes. With Neurala’s system, if you show it a new thing, it says it doesn’t know the SKU. Our system can provide that feedback when it doesn’t recognize, say, a new cereal box, on an as-needed basis.”