Arm on widespread use of properly capable Machine Learning in 2018
With our next AutoSens on the horizon we took the opportunity to catch up with Arm. Arm returns again to support AutoSens, and we are excited to have them at the world-leading vehicle perception conference in Brussels. They gave us some interesting insights into the looming end of Moore’s law, their approach to the open vs closed architecture debate and their excitement about the widespread future use of properly capable Machine Learning.
What are the emerging trends for in-car processing solutions?
One of the emerging trends for in-car processing solutions is the move towards more sophisticated hardware-based processing, with less reliance on large amounts of complex software doing a lot of “heavy lifting” of data. Using the combination of hardware and software in a more efficient manner, playing to the strengths of each and optimising the workload across the system will be critical in advancing the data processing capabilities required for machine-based vehicle control systems.
Will the predicted end of Moore’s law be a hindrance to progress, are there ways around it?
In the context of automated driving systems, the predicted end of Moore’s law is currently unlikely to be a limiting factor; processors are already largely fast and powerful enough; the current focus is on hardware acceleration and more intelligent uses of the sensor data being gathered. Effort expended in processing data that is ultimately irrelevant to the situational awareness of the vehicle, or to any associated control decisions is wasted. Machine Learning algorithms will help reduce the amount of raw data being processed to completion, then subsequently being discarded. Humans do this very well; the key will be having machines approach our capabilities in this regard.
What sort of problems can low dynamic range cause and why is that important for extracting data from video?
Vehicle vision systems must have the ability to acquire low-noise detail in mixed light conditions, without saturation or clipping of either dark or bright areas in the field of view. Low dynamic range images cannot provide this level of detail, and computer vision engines require highly-accurately measured, rather than calculated or inferred data. Providing high dynamic range image data to computer vision engines is a critical part of neural network training and machine learning applications.
How do you feel the industry can help resolve the shortage in skilled people?
To help address the shortage of skilled people, we all need to make a concerted effort to promote the level of cutting-edge engineering and development taking place across the industry to a wide range of young people, starting as early as is practical. Most of what we do is behind the scenes, so people aren’t aware that there are challenging and exciting opportunities available to get involved in the next level of intelligent technology, and that it is probably being developed near you, by people like you, in companies who need more developers.
Watch latest videos from AutoSens on YouTube
► Improving and implementing traditional computer vision algorithms…
► How deep learning affects automotive SOC and system designs
► Coherent Lidar for Autonomous Driving – Radial Velocity on Every Data Point
What is Arm’s approach to the open vs closed architecture debate?
There are demonstrable benefits of both open and closed architectures. But for many applications there are considerations around functional safety and cyber security in the systems incorporating Arm designs that often drive the necessary choice of a closed architecture. There is a balance to be struck between supporting open architecture designs and providing the end-users with the appropriate level of confidence in the design.
Away from processing solutions, what’s the most exciting technology change in industry you expect to see in 2018?
The most exciting technology change in the industry that we expect to see in 2018 is the widespread use of properly capable Machine Learning. This has the potential to enable huge steps in the capabilities of data processing systems and the tools used to calibrate and tune them. Once we move from the “brute-force” approach into more intelligent mechanisms for processing very large datasets, rapid development and departures into new applications on smaller devices at the edge become achievable.
What else are you most looking forward to at AutoSens in 2018?
AutoSens is a great opportunity for learning more about the technology roadmaps of automated driving system and component developers, sharing ideas and collaborating on the next generation of technology.
Come and meet Arm at AutoSens Brussels (17-20 September 2018 at AutoWorld Brussels) – ticket available here >>