Standardized Neural Networks Could Help Machine Learning

The ever increasing volume of data captured by technology continues to transform the nature and direction of business activities across the globe and the automotive sector is no different.
This ongoing trend means that data-led approaches are an increasingly common influence on the design, as well as the development of a wide range of features and value-added services across the car industry. The growing use of data analysis methods and machine learning algorithms is also helping to underpin the creation of next generation autonomous vehicles.
Data led development
One company that has proved particularly willing to embrace the opportunities afforded by Big Data is the German automotive behemoth BMW, which has adopted a data-driven approach to a comprehensive range of activities – from the initial design, engineering and manufacture of vehicles to sales, after-sales and customer support. In doing so, it has entered into collaborative partnerships with a number of companies in an effort to cultivate the synergies necessary to negotiate the increasingly complex technological landscape. For example, it has teamed up with Intel, which itself had recently bought out Mobileye, a front-runner in computer vision technology, to work on the development of level three and four autonomous vehicles. The German powerhouse has also entered into agreements with trailblazers in the artificial intelligence (AI) sector, including IBM via the integration of Big Blue’s Watson AI-led cognitive computing business platform in prototype i8 hybrid vehicles to deepen understanding of how drivers can interact more seamlessly with on-board systems, handle communications with other driverless vehicles and self-diagnose faults.
Elsewhere, electric vehicle firm Tesla is continuing to develop a range of data-driven services. Since its inception, Elon Musk’s disruptive venture has made it company policy to capture and store as much customer data as possible before transferring it to cloud servers, where it is analysed using state of the art software and AI algorithms. The knowledge gained as part of this process has enabled the company to offer a constant stream of over-the-air (OTA) functional upgrades, which can also be sold per hour or mile.
Here and now
Another company that is very much at the forefront of automotive sector efforts to use data analysis and interpretation in the design and development of autonomous cars is the Amsterdam, Netherlands-based outfit Here, which is majority-owned by a consortium of German automotive companies, including Audi, BMW, and Daimler. One of the company’s key technologies in this field is the Here HD Edge Perception stack, which employs deep convolutional networks to detect features in video streams in real time. As Stephen O’Hara, principal engineer, highly automated driving at Here, explains this involves what he describes as a ‘data-driven process’ where sample images are collected using vehicle-mounted sensor systems, before being annotated by dedicated labeling teams- and then used to train neural networks to detect relevant features in the images.
“Examples include detecting signs, poles, lane markings, vehicles, road-side barriers and so on. In the case of edge perception, neural networks must run on low wattage computing devices while maintaining high throughput,” says O’Hara.
On the road to full autonomy
In O’Hara’s view, data driven learning will very much be required to increase and improve autonomy and he highlights the fact that the breakthrough of deep learning was the “move away from hand crafted, low-level descriptors used for feature detection towards having the network learn how to generate low-level descriptors best adapted for a given task”.
“We now need to take general operating and safety principles and encode them into the learning process,” he says. “To some degree, developers are using simulators for safety training but simulators are not yet as realistic as they need to be for visual perception.”
Looking ahead, O’Hara predicts that one of the key innovations in data driven services over the next few years will relate to the ongoing trend towards standardizing the representation of neural network architectures, for example, via ONNX, the Open Neural Network eXchange which he believes allows developers to more easily separate training frameworks, such as TensorFlow, PyTorch and MxNet, from deployment and optimization.
“Custom hardware for accelerating neural nets or improving their run-time efficiency, that is their power usage, is trending up, and with standardized formats such as ONNX, compilation tools can be developed to target a single model to various hardware deployments,” he adds.