Brodmann17: Automatically training mobility-grade neural networks on Google Cloud

About Brodmann17

Brodmann17 develops perception software for automated driving. The solution exceeds state-of-the-art accuracy while consuming a fraction of computing power, enabling the transition of AI capabilities from premium to mass-market. The technology is hardware-agnostic, allowing Tier-1 suppliers and automakers to combine any visual sensor with any processor ranging from L1 to L4 automation.

Industries: Technology
Location: Israel

Brodmann17 built an automatic mobility-grade, neural networks training platform, which scales on demand and automates the process of neural networks creation for different products.

Google Cloud Results

  • Helps ensure the best algorithms and neural networks while reducing infrastructure costs and time
  • Automates neural networks creation to improve scalability and help ensure the highest quality for all customers
  • Empowers engineers to spend more time working on core objectives for perception software without worrying about infrastructure management

Cuts infrastructure costs by 65%

Once a science fiction concept, artificial intelligence (AI) is an increasingly large part of everyday life. Made up of neural networks that are trained from huge datasets, AI can be used for a wide variety of tasks, from voice recognition to automated chatbots to self-driving vehicles. But vision perception has proved a challenge as it needs more compute and processing power than most other tasks. Rather than wait for the next generation of processors, Israel-based Brodmann17 aims to bring powerful vision perception capabilities to today's market by improving the efficiency of the neural networks for mobility that make up its AI platform.

The company's founders developed a highly efficient, proprietary algorithm that allowed its neural networks to operate at a fraction of the compute power required for its competitors. "In most benchmarks, we require about 5 percent of the resources of other neural networks. In some cases, we were as much as 40 times as efficient," says Brodmann17 CTO and co-founder Amir Alush. "That allows us to run powerful neural networks on standard processors which you find in everyday devices."

"We wanted to keep our automated training platform flexible and scalable with our existing cloud, but we needed it to be more effective. After testing all the major cloud providers, we chose Google Cloud Platform due to its lower costs, streamlined DevOps, and superior performance in benchmark tests."

Amir Alush, CTO and Co-founder, Brodmann17

The company has already worked with major automotive industry companies, which are keen to exploit AI now. Brodmann17 used the cloud to train its perception models, but as the company grew it realized it needed a new infrastructure to cope with the extra demand. It turned to Google Cloud Platform (GCP).

"We wanted to keep our automated training platform flexible and scalable with our existing cloud, but we needed it to be more effective," says Amir. "After testing all the major cloud providers, we chose Google Cloud Platform due to its lower costs, streamlined DevOps, and superior performance in benchmark tests."

Compute Engine for cost-effective high performance

While Brodmann17's neural networks require minimal power at the inference, the training process and the models that underpin them require vast resources. Looking for a new, sustainable infrastructure that would minimize expenses without affecting the quality of the models, the company migrated its infrastructure to Google Cloud.

Initially, the process of training neural networks was manual. The engineers would set up an instance, load the data by hand, copy across the training code, run it, and then output the results. As the company took on more products and more complex tasks, the training process began to take up increasing amounts of its time. As a fast-growing startup, Brodmann17 put great importance on efficient allocation of resources and began developing an automated mobility-grade training process.

"The key for us is how well Compute Engine graphical processing units perform. When we're automatically training several networks together, latency becomes a critical issue. With Compute Engine, we're able to deliver the best products for our customers."

Amir Alush, CTO and Co-founder, Brodmann17

After assessing its options and choosing GCP, Brodmann17 began preparations for the migration, shifting its training platform over to GCP, ably assisted by the Google Cloud team in Israel. "It felt very natural and easy to work with," says Amir. The company held its training data in highly secure Cloud Storage and then the proprietary training platform would run on Compute Engine instances.

"The key for us is how well Compute Engine graphical processing units perform," says Amir. "When we're automatically training several networks together, latency becomes a critical issue. With Compute Engine, we're able to deliver the best products for our customers."

Using monitoring data and logs from Stackdriver, Brodmann17 developed its own algorithms to determine which combination of resources was best for different tasks. Once it was comfortable with the new infrastructure, the company began to automate and synchronize its DevOps tasks with Google Kubernetes Engine (GKE) and App Engine, allowing it to train far more neural networks than it could with the previous manual solution.

"With Google Cloud Platform, we've built a scalable, flexible, and cost-effective infrastructure to train our neural networks. It lets us build more and better neural networks for our clients."

Amir Alush, CTO and Co-founder, Brodmann17

Automated DevOps, more focused engineers with Google Cloud

Thanks to the company's savvy use of resources, Brodmann17 has reduced its infrastructure costs by 65 percent according to Amir. "With Google Cloud Platform, we've built a scalable, flexible, and cost-effective infrastructure to train our neural networks," he says. "It lets us build more and better neural networks for our clients."

Meanwhile, the ability to automate its infrastructure management with GKE and App Engine enabled Brodmann17 to work at a scale that was impossible with a manual setup. Within a year of migrating to GCP, the company had used millions of hours of compute instances for training its neural network. The automation has allowed the company to run hundreds of neural networks simultaneously, which would have been impossible with its previous setup.

Since the migration, Brodmann17 has been working with Google Cloud to see how it can expose more of its training infrastructure to third parties. This would help Brodmann17 reach newer, bigger clients that need to bring the training models into their own servers. The latest iteration of the automated training platform has made significant efficiency gains by combining App Engine with logs and monitoring data from Stackdriver. For Brodmann17 and Google Cloud, the work to bring the next generation into the present never stops.

About Brodmann17

Brodmann17 develops perception software for automated driving. The solution exceeds state-of-the-art accuracy while consuming a fraction of computing power, enabling the transition of AI capabilities from premium to mass-market. The technology is hardware-agnostic, allowing Tier-1 suppliers and automakers to combine any visual sensor with any processor ranging from L1 to L4 automation.

Industries: Technology
Location: Israel