Artificial Intelligence In Autonomous Vehicles
Artificial intelligence is often considered the key to autonomy for self-driving cars. But why is that? The human being is remarkable in that it can drive vehicles at high speed and in urban areas with many things happening on the road and to the side – and yet a person can, usually, manage all these visual and aural cues without causing an accident.
For all these inputs, such as traffic lights, pedestrians, other cars on the road, and roadworks, and challenges such as rain, fog and snow, the vehicle needs to replicate the ability of the human brain to sense what is happening. It needs to anticipate and predict what will happen next, and then recommend a course of action. The AI has to decide on whether to slow the car down by applying the brake or to drive carefully to avoid an obstacle.
The AI doesn’t work in isolation. Artificial intelligence advances rely on autonomous hardware. In turn, that hardware relies on autonomous software. The three, together, deliver truly autonomous vehicles.
As the growth in capability of trained neural networks has increased there have been many breakthroughs for artificial intelligence. The self-driving car is, therefore, becoming a reality.
The car as a data centre on wheels
AI is now a fundamental part of our world. Though we may not be aware of it, artificial intelligence is involved whenever we search for a word on the web, when we make a financial transaction or when we have a filter applied to a picture on our smartphones.
Through neural networks, AI has been putting the smart in smartphone, and now it’s putting the smartness in smart cars. Neural networks, which require considerable compute power to be effective, are being run ever more efficiently, due to advances in hardware acceleration. Originally running on CPUs, they initially received a massive speed boost thanks to the power of graphics processing units (GPU), which through their inherently massively parallel compute engines, proved ideal for running compute tasks.
However, as the name suggests, neural network accelerators are designed specifically for these tasks. As such, they are far more power and area efficient than even GPUs, making them orders of magnitude more effective at running neural networks for inferencing on the edge.
The self-driving car is becoming a reality.
This efficiency means it’s now possible to run the required AI models directly on an edge device, (such as in a vehicle), without relying on cloud computing, which would entail the round-trip time of contacting a data centre. It means we now get the power and predictability of running neural networks locally, something that only a few years ago would have been the exclusive domain of the data centre – but now is possible on the data centre on wheels.
By running neural networks locally, all benefits of the autonomous car, from functional safety to multi-core flexibility, can run quicker and faster, and all with low power consumption.