EML Blog #6

The type of machine learning I chose to look into was machine learning for automated driving of cars. This has been introduced the the commercial market and has been picked up by manufacturers/developers (most notably Google) but it is not yet available to the general public. What is at stake here is significant, the obvious repercussions of poorly implemented products being made available to the general public being death and injury to drivers and pedestrians. And what has held it up for being fully available is that exact risk (http://spectrum.ieee.org/cars-that-think/transportation/advanced-cars/deep-learning-makes-driverless-cars-better-at-spotting-pedestrians).

While the algorithms used in Google’s cars have been notably successful, these cars needed and still need deep learning to take place on their programs in order to ensure 100% pedestrian safety. As mentioned in the article, this deep learning practice also is benefiting crash avoidance systems in cars that are currently available to the public in a retail sense. The algorithm in the article can currently read pedestrians at a rate of 2–4 frames per second which is still not fast enough to ensure the safety of everyone involved to implement self-driving cars, but has done great things for crash avoidance systems in cars currently available on the retail market.

It seems that is the crux of where this development is. Further learning and fixes need to be made for the algorithms for Google to take the next step in providing these cars on the retail commercial market (although as I mentioned it is available from an institutional perspective) but it has been implemented fairly successfully for cars being bought in the form of crash-avoidance systems.

Here is a similar article on Google’s version of the algorithm: http://spectrum.ieee.org/cars-that-think/transportation/self-driving/new-pedestrian-detector-from-google-could-make-selfdriving-cars-cheaper