A Blog by Jonathan Low

 

Jul 16, 2018

Infinite Edge Cases: Will 'Driverless' Cars Always Need Human Involvement?

There will probably always be exceptional, so-called 'edge' cases.

Which means human involvement to figure them out will continue for the foreseeable future. JL

John Naughton reports in The Guardian, photo by James Sensuan:

Accidents are what engineers call “edge cases” because the cars’ software encountered scenes or objects that it didn’t recognise for the simple reason that they had never appeared in the software’s training datasets. Given that the real world is full of things that cars’ sensors and software have never seen before, these edge cases will continue. Chance and human nature will ensure that the number of conceivable edge cases is very large (if not infinite).
In 1979, Douglas Hofstadter, an American cognitive scientist, formulated a useful general rule that applies to all complex tasks. Hofstadter’s law says that “It always takes longer than you expect, even when you take into account Hofstadter’s law”. It may not have the epistemological status of Newton’s first law, but it is “good enough for government work”, as the celebrated computer scientist Roger Needham used to say.
Faced with this assertion, readers of Wired magazine, visitors to Gizmodo or followers of Rory Cellan-Jones, the BBC’s sainted technology correspondent, will retort that while Hofstadter’s law may apply to mundane activities such as building a third runway at Heathrow, it most definitely does not apply to digital technology, where miracles are routinely delivered at the speed of light. Think of the astonishing advances in machine learning, for example, or the sophistication of smartphones. Or think of the self-driving car, an idea that seemed preposterous only 15 years ago and yet is already a reality on the highways of a number of US states. Surely these and other achievements of digital technology took less time than we thought?

Well, yes and no. Having ideas and building digital prototypes to implement them can indeed happen very quickly. Incorporating them into everyday life, however, is not. Cue Rodney Brooks, a world-leading robotics expert and one of the wisest commentators on digital technology known to this columnist. “Having ideas is easy,” he writes. “Turning them into reality is hard. Turning them into being deployed at scale is even harder. And in evaluating the likelihood of success at that I think it is possible to sort technology and technology deployment ideas into a spectrum running from relatively easier to very hard.”
On that spectrum, changing anything that involves just software – programs – is relatively easy. Once hardware – physical kit – is involved it very quickly gets much harder. Most of the world’s factories, for example, are controlled by programmable logic controllers or PLCs. “When you want to change information flow, or control flow, in most factories around the world,” writes Brooks, “it takes weeks of consultants figuring out what is there, designing new reconfigurations and then teams of tradespeople to rewire and reconfigure hardware. One of the major manufacturers of this equipment recently told me that they aim for three software upgrades every 20 years.”And if a new technology is going to have critical interactions with humans or public institutions, then Hofstadter’s law will apply in spades. Which brings us neatly to self-driving cars, those poster-children for digital innovation.
If you believe the hype, you’ll be expecting to see them on our roads any day now. Dream on. That doesn’t mean that the technology isn’t amazing. It is. And the vehicles have an impressive safety record compared to the carnage wrought by human-driven vehicles. Their excellent performance is largely enabled by machine-learning algorithms that have been trained on torrents of data about roads, intersections, street furniture, etc. But sometimes there have been fatalities. In 2016, for example, a self-driving Tesla Model S drove into the back of a white tractor-trailer; in March this year, a self-driving Uber car killed a woman pushing a bicycle; and so on.
These accidents are what engineers call “edge cases” because the cars’ software encountered scenes or objects that it didn’t recognise for the simple reason that they had never appeared in the software’s training datasets. Given that the real world is full of things that cars’ sensors and software have never seen before, these edge cases (and their attendant fatalities) will inevitably continue.
So what will happen with driverless cars? The answer is that they will eventually arrive, initially licensed for carefully circumscribed uses, possibly on designated urban streets and separated from humans driving old-style automobiles or mounted on bicycles, scooters, even pogo sticks. In that sense, the cars will be more like the driverless trains that take you to the departure gates in airports such as Stansted or Singapore’s Changi. That doesn’t mean that driverless cars won’t be useful, maybe even welcome. But if you’re wondering why it might be 2050 before they’re a common sight on every street, then there are two obvious answers. One is that chance and human nature will ensure that the number of conceivable edge cases is very large (if not infinite). The other is that Hofstadter’s law will apply to this technology, just as it applies to everything else.

0 comments:

Post a Comment