AI chip manufacturer Hailo has recognized the risks with latency issues and autonomous vehicles. As an example, imagine a self-driving car making its way down a street when a ball rolls in front of it. Current technology may enable these vehicles to recognize the ball itself- but how much time will it take for the autonomous car to deduce that a child may soon be in the car’s path to retrieve that ball?
With autonomous vehicles there can be no time hesitation for latency, data must be processed swiftly and efficiently. Recognizing such latency issues, Hailo has designed a specialized deep learning processor that can adequately analyze reams of data to optimize a vehicle’s perception of its environment. To learn more, Digital Journal spoke with Orr Danon, CEO and Co-Founder of Hailo.
Digital Journal: How important will autonomous cars be for the automotive sector?
Orr Danon: We as a society are on a journey towards reaching fully autonomous driving, yet we are still quite a distance from the finish line. The fact that true autonomy is at least a decade away by most accounts means that we should focus now on benefiting from all the autonomous building block technologies that are being developed in the interim on the road to autonomy.
For example, there is much room for improvement in the performance of sensors in advance driver-assistance systems (ADAS). Although ADAS are currently unable to engage with the vehicle systems themselves (this would make them autonomous), the systems of today are able to provide warnings and assistance to drivers. This includes lane departure warnings (including for blind spots), collision avoidance systems, rear-collision warnings, parking assist technologies, adaptive cruise control, and more. All these advances already help drivers by reducing collisions and making driving a less stressful, safer experience. It’s these technologies that will one day pave the way for cars to reach full autonomy.
DJ: What are the limitations with most autonomous car technology?
Danon: One of the biggest challenges in the industry’s quest to develop autonomous vehicles is that it still relies on decades-old computer architecture for which performing deep learning tasks is immensely challenging. While autonomous technologies currently do exist, like self-driving buses that operate in controlled environments such as college campuses, these systems still call for more efficient processing in order to better interpret their surroundings and avoid accidents.
Autonomous vehicles today either ferry data to faraway data centers for analysis – which causes untenable latency issues where every split-second counts – or they carry around bulky data centers in their trunks to process sensory perception. The latter adds dangerous weight to vehicles and requires enormous power consumption that causes untenable heat dissipation issues which require even more bulky cooling systems – in short, a non-starter.
Self-driving vehicles must be able to address critical mission tasks by processing data quickly and accurately in order to understand the vehicle’s environment – including accurately identifying pedestrians, vehicles, and road conditions – to make almost instantaneous decisions necessary for self-driving.
However, current solutions are not suited for these complicated challenges. Such limitations are putting the brakes on the progress of self-driving cars. Imagine an autonomous vehicle making its way down a street when a ball rolls in front of it. How long will it take for the car to recognize the ball itself and then deduce that a child may also soon be in the car’s path to retrieve this ball? That’s quite a sophisticated task which takes enormous computing power on the device itself and there is no time for latency: data must be processed swiftly and efficiently through high performance computing in the vehicle itself.
DJ: How can deep learning overcome these challenges?
Danon: Deep learning algorithms provide the best solution for many of the challenges I mention above. For instance, deep learning can enable cameras to perform tasks such as semantic segmentation and object detection more efficiently. When being given enough compute resources to run at high resolution and high speed, deep learning algorithms enable vehicles to understand their environment in real time at a reasonable distance. Vehicles able to perform image processing more efficiently are able to make better instantaneous decisions. This means both ADAS and autonomous technologies can greatly benefit from effective deep learning algorithms.
DJ: How was Hailo developed?
Danon: Hailo was founded in 2017. I was working with my co-founder Rami Feig, who I met while serving in the Israeli Defense Forces’ elite intelligence unit, on cybersecurity solutions for IoT. The then relatively unknown term “deep learning” kept popping up throughout our research. Eventually, we decided to bring together a team of experts, including Avi Baum, an expert in IoT and machine learning, who was then serving as the CTO of Texas Instruments’ IoT division, and Hadar Zeitlin, a smart-mobility guru, to develop a new deep learning chip from the ground up – one that addressed the shortcomings of aging processing architecture to enable smart devices to operate more effectively and efficiently at the edge. After Rami’s unfortunate passing, Avi, Hadar, and I helped realize his vision of creating Hailo’s groundbreaking processor. The chip is built with an innovative architecture that enables edge devices to run sophisticated deep learning applications that could previously run only on the cloud. To date, we have raised $28M and officially launched our product this year.
DJ: What were the main technological challenges with the development?
Danon: We realized that industries relying on decades-old computer architecture face processing bottlenecks and struggle deploying this technology at scale. Today, established semiconductor companies still use designs based on the same underlying concept of the 70-year old Von-Neuman architecture. A full high-definition camera operating in real-time in a vehicle, for instance, requires the full capacity of a data center scale GPU, consumes dozens of Watts per sensor and as a result requires cumbersome active cooling systems.
Our key technical challenge was to reimagine the industry’s typical data processing and to provide the market with a chip that does not need to compromise on performance or power consumption. We have the flexibility to rethink how current embedded systems –and future ones – need to process unstructured data, and so we developed a processor architecture to execute complex heavy structures rather than rules-based ones. We found that domain-specific processors built with deep learning architecture reduce the power and costs needed to analyze data, enabling devices to operate at the edge, with lower latency, higher security, more privacy, lower costs, and higher reliability.
Hailo is now sampling its breakthrough chip with select partners across multiple industries, including automotive, smart cities, and smart homes. Hailo is working with leading OEMs and tier-1 automotive companies in fields such as advanced driver-assistance systems (ADAS). These partners include three of the world’s largest automakers.
DJ: What are your predictions for the next couple of years?
Danon: AI chips will revolutionize multiple industries. As with autonomous vehicles, where processors like Hailo’s will improve cars’ abilities to efficiently process visual data, AI chips will also revolutionize fields such smart cities, improving public safety by assisting with crucial tasks such as locating missing persons or finding stolen vehicles. Such processors will also enable real-time traffic monitoring, increased traffic safety, and improved transportation flow monitoring, thereby cutting down on traffic snarls and pollution. For smart home devices like virtual assistants, chips like Hailo’s can process data at the source, translating into more privacy, more reliability, and faster response times. In short, AI chips are empowering a whole new era of computing at the edge and this is only the beginning.
