I was driving my car the other day and I heard a commercial for a car that would make driving safer by using Artificial Intelligence. In the commercial the voice claimed that while other brands are endeavouring to “chase the future” this particular brand was creating and even anticipating the future.
I am not sure how much marketing and how much truth was in their statements but I was anyhow impressed that a car manufacturer has decided to use AI as a way to attract customers attention and to make it a selling point. It is, irrespectively of truthfulness, a bold step to assume that people are willing to release their grip on driving the car and accepting AI taking over.
AI is indeed a vital component in self driving cars and not because they need to be smart, intelligent, to self-drive (at least not just for that) but to “decrease” cost! If you are smart you consume much less resources, including energy. A professional basketball player uses a fraction of the efforts I would use to aim at the “basket” (and with much better results, of course). Our brain is a clear example: the better you are at something, be it doing calculus or painting a wall, the less resources are used by the brain. The miracle is achieved through specialisation of neural circuits (these are formed dynamically throughout our life). A different word for specialisation is “learning”.
A team of researchers at MIT has developed a programmable nanophotonic processor that can support deep learning and that has been specifically designed for application in a car environment to serve as the car brain. The use of photons, rather than electrons, make this circuit particularly efficient from the point of view of energy requirements.
The idea is to use “beams of photons” and to steer the way the merge (input signal). This creates interference bands that are the output signal. In practice by modulating photon beams one can “compute”. This is similar to what happens in our brain circuitry, where signals “interfere” with the neurones potential causing some to fire. Notice that the interference bands created can be mapped onto matrices multiplication, something that is required in deep learning algorithms and that is processing intensive. The protonic processor developed at the MIT can implement these matrices multiplication very effectively.
There will still be quite some work required to scale up the processor and make it applicable to AI problems but the approach proposed is interesting.
The fact that the researchers are explicitly indicating self driving cars as ideal applications provides a specific direction for the future work.