Blog

March 12, 2018

Neurala Makes Machines Intelligent, One Brain at a Time

A lot of startups incorporate machine learning into their solutions, but far fewer attempt to fundamentally advance the technology in-house. Such a strategy requires founders to possess not only deep technical expertise but also the ability to maintain a business mindset while pursuing research objectives.

That can be a difficult balance to strike, particularly for founders coming from research labs, but the people that can pull it off gain a distinct advantage by offering products that leverage truly innovative, proprietary technologies.

The founders of artificial intelligence startup Neurala are attempting to tow this line, and although they see R&D as a key differentiator for the company moving forward, they’ve already proven there’s a business application for their ideas. In fact, there’s a lot of applications.

The company’s deep learning system, the Neurala Brain, is being deployed in drones, toys, body-worn cameras, and more to make the objects around us more intelligent.

What really makes the solution stand out, however, is a proprietary technique that allows it to learn new objects after it's been trained. And the system's ability to run on low-cost processors without relying on the cloud or internet access puts it in a good position to play a role in the race to make AI ubiquitous.

Now, roughly a year removed from a $14 million Series A funding round, the founding team’s ability to scale will determine the extent of Neurala’s success.

The Neurala Team

The company’s co-founders, Heather Ames, Massimiliano “Max” Versace, and Anatoly Gorshechnikov, met while pursuing their PhDs in cognitive and neural systems at Boston University in the early 2000s.

Neurala's founding team from left to right: Massimiliano Versace, Heather Ames, and Anatoly Gorshechnikov

At the time, the three founders shared server space to run brain models that often took days to produce results. If they wanted to tweak even a small parameter, they’d have to start the entire model over again and reset the clock on results.

One day at a coffee shop Versace predicted to Gorshechnikov, “Some day we’ll be able to run these models on a cell phone.” Gorshechnikov promptly spit out his coffee.

So began conversations about running their models on different types of hardware, culminating in the realization that graphical processing units (GPUs) and other forms of customized hardware would be a far more efficient and effective way of running artificial neural networks.

The researchers knew they were on to something, so in 2006 they founded Neurala and filed a patent to process artificial neural networks on GPUs.

“We were way ahead of the curve on that,” Ames, who’s now Neurala’s chief operating officer, said. “At that point, we were still in graduate school, so admittedly we didn’t do much with Neurala in those days, but we kept working in the background on that general concept because we knew we were really early in the marketplace.”

As the founders continued their research at BU, they took part in several ambitious government projects that forced participants to push the limits of AI systems and rethink traditional computing methods.

Members of DARPA’s ongoing SyNAPSE project, for instance, are attempting to create computing architectures that replicate the structure and functioning of neurons in a portion of mammal brains. Neurala’s team built software to run on these massively complex systems in partnership with Hewlett-Packard (as part of the project, IBM would later produce a computer chip with more than five billion transistors).

When officials from NASA approached the founders in 2010, they wanted to use AI on the Mars Rover to help the robot explore the red planet autonomously.

“[NASA]’s ultimate objective was determining how [autonomous systems] could run in GPS-denied environments to be able to navigate and discover things,” Ames said. “To do that second part the system needs to be able to stop when it sees something interesting and gather more information, but because it’s on Mars it has to make those decisions on its own.”

Each of these projects helped the founders get their deep learning system closer to feasibility in business environments. Finally in 2013 Neurala came out of stealth mode and joined the Techstars program to help bring its standalone deep learning system to planet earth.

Four years later, Neurala is a carefully scaling company with the ability to bring intelligence to objects as diverse as robots, self-driving cars, drones and children’s toys.

“Maybe we have some attention issues,” Ames joked about the company’s broad potential customer base. “It’s just because the area is so new people are still exploring it. We’re a long shot paving the way in this industry trying to figure out what’s possible.”

The Neurala Solution

The Neurala Brain is a highly efficient system that can run locally on anything from NVIDIA, ARM and Intel processors to cheaper, single-board computers like the Raspberry Pi 2 and the Pine64 family of computers.

After training Neurala’s “Brains for Bots” SDK with only a handful of examples, the system gives anything with a camera the ability to find, recognize and track objects in real-time from images. That means drones can identify and follow their owners mid-flight, surveillance cameras can find guns as they appear in video feeds and cars can gain a deeper understanding of their surroundings by recognizing bikers and pedestrians as more than just obstacles.

Customers purchasing Neurala’s SDK can train it themselves without prior knowledge of building deep learning systems. The company also works with customers to train systems by either collecting labeled data sets and customizing a system under its “Brain Builder" program or by scanning customers’ data directly.

“There’s a lot of back and forth,” Ames explained. “AI is fairly new and it’s changing the face of technology. We educate our customers on the kinds of data that works and they educate us on their use cases.”

The Neurala Brain’s ability to run locally means machines don’t need to be connected to the internet (or any part of the outside world for that matter). This feature gains particular importance when you’re putting cameras in children’s toys. And the ability to learn new things quickly is important because, as Gorshechnikov recently exclaimed to a roomful of stern AI scientists, “Has anyone ever seen a child who looks at a toy for more than five minutes?”

A number of the features mentioned above illustrate the latest progress in deep learning technology, but Neurala’s announcement in May of last year might represent its biggest breakthrough yet.

A fundamental problem with deep learning (and a cause for some to question the efficacy of the field going forward) has been that these systems stop learning new things once they’ve been trained. Teaching a system anything post-training has required users to completely reset the system and start over (this problem is referred to as catastrophic forgetting by researchers).

Neurala began incorporating a technique to overcome this problem into its solution in the third quarter of 2017. The patent-pending technique, which uses what the company refers to as Lifelong Deep Neural Networks (Lifelong-DNNs), allows users to add objects to the system’s knowledge base in seconds using only their cell phones without retraining the entire network.

“[The Lifelong-DNN] approach is the enabler that automotive companies, consumer electronics companies and others have needed to make deep learning useful for their customers,” Versace, who serves as Neurala’s CEO, said in the press release announcing the breakthrough. “The ability to learn on the fly and at the edge means that the Neurala approach enables learning directly on the device, without all the drawbacks of cloud learning… Most importantly, it will unlock the development of a sea of cloud-less applications.”

The founders are staying tight-lipped about the technology behind Lifelong-DNNs, but a key enabler for this capability is that the system doesn’t use back propagation, a popular method for training artificial neural networks that diverges from the learning mechanisms of biological systems.

Instead, Gorshechnikov has said the system partly uses “Hebbian-like learning” methods, referring to the Hebbian theory of neuroscience that describes how biological brains learn things.

“We emulate different aspects of cortical and subcortical processing in software,” the founders said in a statement. “The full details of the tech are patent-pending, and we can't discuss them now, but we believe they will be as powerful and obvious as our original patent centered around using GPUs for AI.”

Neurala’s Path Forward

People often say deep learning systems, with their layers of connected artificial neurons, mimic the structure of biological brains. The truth is that no system comes close to matching the complex interactions occurring between the billions of neurons contained in the human brain.

Still, Neurala’s founders seem to have taken biological emulation a step further by considering the designs of the brain models they developed at BU to develop their system.

“Our neuroscience backgrounds help guide our thinking a lot,” Ames said. “When we run into tough problems we go back to our initial training and say, ‘Well how do humans or rats do this?’ We’re constantly asking those questions, particularly in our R&D group. They don’t just look for the simplest shortcut, but instead they ask ‘What would be the more biologically faithful solution?’ And then ‘How can we get to the best solution for our customers?’”

Hearing the founders discuss Neurala’s research strategy makes it clear they’ve moved on from their days in academia, where they’d study whatever problems seemed most interesting, and are now focused on creating a successful company.

“A lot of our peer companies that start from research departments don’t have an eye toward learning how to productize and apply this technology,” Ames said. “We don’t want to just be a handful of PhDs with a great idea. The co-founders are constantly looking at things from a biological standpoint, but our strategy is to prioritize the research efforts that have a close proximity to customer needs.”

Neurala already counts big-time customers like Motorola and Parrot Drones as partners and drone companies make up the majority of their current customer base.

Of course, with the Neurala Brain’s ability to run on so many different types of hardware, the company is also working to bring intelligence to the never-ending mass of electronic toys rolling off assembly lines around the globe.

Such is the explosively-changing nature of AI advancements today. The technology in yesterday’s Mars Rover will be in the Barbies and Elmos of tomorrow. In such a climate, today’s pioneers can get left behind if they don’t keep moving in the right direction.

Fortunately for Neurala, innovation is in their DNA.


Zach Winn is a contributor to VentureFizz. Follow Zach on Twitter: @ZachinBoston.

Images courtesy of Neurala.

Our current openings