Recent advancements in artificial neural networks (ANN) and so-called deep learning are accelerating the reality of self-driving vehicles faster than was originally expected as hardware and software vendors are taking the lead in pushing the technology to enable autonomous vehicles forward.
Deep learning, also known as machine learning, has been a concept in place since the early 1980s but only recently has technology advanced to a point where it has become a feasible reality. The idea of deep learning is to attempt to artificially emulate the functionality of the human brain via hardware and software. An ANN will continuously learn and will base its ability to recognize the surroundings on a deep learning phase based on real examples of sounds, images and input from other senses.
The ANN is defined and trained with several images—with features such as edges, colors and shapes—and is able to associate the input to what statistically is believed to be the target. The advantage here is that the artificial network is able to evaluate the context of the image it is seeing rather than just the recognition of an object. This is seen as a key step toward the enablement of autonomous driving, according to Luca De Ambroggi, principal analyst for automotive semiconductors at IHS.
Because this allows for a well-trained ANN to evaluate in real-time objects and also the context in which the objects are seen, a vehicle, for example, could not only detect a person crossing a street but also determine if that person is distracted (on a cell phone or not looking at oncoming vehicles). Not surprisingly, this requires an enormous amount of processing power.
Error Rate Records
Recent breakthroughs in ANN and deep learning concepts have been achieved to the point where the error rate for machine vision has surpassed that of a human being’s capabilities, which is set at about 5% error rate, De Ambroggi says. In 2011, the error rate was 25% and it fell to below 16% in 2012. However, recently both Microsoft and Google have set error rate records of 4.9% and 4.8%, respectively. Chinese Google competitor, Baidu, recently announced a 6% error level.
This record rate was due to new architectures being created using multiple graphics processing unit (GPU) cores rather than traditional CPUs. “Semiconductor vendors are contributing in two ways through software and algorithms supported by extremely powerful parallel processing,” De Ambroggi says. “For the learning phase, this also means storing a lot of data somewhere, such as through storage houses from Google or Baidu, where that data will be able to feed the system.”
Google is investing heavily in autonomous vehicles and already has an active system integrated in its self-driving vehicle showing how deep learning is able to detect pedestrians in different situations, De Ambroggi says.
“If implemented right, this is a breakthrough for autonomous driving,” De Ambroggi says. “Because machines will be able to get closer to what real driver perception is.”
Nvidia Drives Forward
Graphics chip leader Nvidia has announced its plan to enable ANN with its silicon control units and is already working with Audi’s zFas platform for autonomous driving. IHS believes that Nvidia has also started to ship its development platform, dubbed DrivePX, to major “premium” automotive OEMs.
However, Nvidia, while it is well positioned De Ambroggi says, is not the only chipmaker active in the enablement of deep learning. MobilEye, leader in car image processing solutions, has also announced plans to enable ANN with their silicon units.
ARM and IBM are very active in ANN, although they are not specifically focusing on automotive yet. Intel, Renesas, Texas Instruments, Freescale/NXP and STMicroelectronics are also in the mix. And field programmable gate array (FPGA) vendors Xilinx and Altera (now Intel) are expected to play here too, thanks to the high processing parallelism of their solutions.
“All of the major processor suppliers will be active in ANN and deep learning and there is a good opportunity for new startups to take advantage of the market especially in neural network algorithms,” De Ambroggi says.
Hurdles to Overcome
The recent progress in deep learning has accelerated faster than anyone expected and will no doubt hasten the reality of self-driving vehicles. However, even with the rapid availability of “self-determining” technologies, liability aspects are likely to slow the whole process, not considering the missing legislation that should allow and rule on self-driving vehicles. OEMs will be under much greater scrutiny if machines are to take over the role of humans in vehicle control, De Ambroggi says.
Still, De Ambroggi says there are other applications where deep learning can be used for immediate advantage to drivers where legal ramifications will be less constrained such as in-vehicle human-machine interface (HMI), like for voice/gesture recognition, or drowsiness detection.
Questions or comments on this story? Contact [email protected]