Slashdot Mirror
Google's Driverless Cars Now Rolling In the Heart of Texas
MarkWhittington notes that, as reported by The Wall Street Journal, Google has started testing its self-driving cars in Austin. These driverless cars, loaded with the sensors, GPS transponders, and cameras, are now in service in "an area northeast and north of downtown Austin. The purpose of the test drives is to see if the car's software works in driving conditions outside of California and to develop a detailed map of Austin city streets. Each self-driving car has two human drivers ready to assume manual control if something goes wrong."
How many drivers does it take to drive a driverless car?
It shouldn't take long for some of the inhabitants to consider this to be the tip of the Obama Administration spear to take over Texas so they can remove their guns, impose environmental regulations, force money to be spent on education. And this right after the Jade Helm 15 exercises. They are probably Islamic driverless cars.
Wow, checking if something works outside of California. What an idea!
Wake me up when they test winter driving in upper New Hampshire.
OK here's the thing with this generation of driverless cars- their motion is governed by neural nets. I am going to assume that everyone here is familiar with this programming paradigm. If not, the Wikipedia entry on it is adequate.
While in the end NN are just another form of Turning machine, currently no one can divine the algorithm of a trained neural net well enough to express it in IF THEN ELSE WHILE form.
That means given a trained NN which is 100% correct 100% of the time , no could write an imperative or procedural (broadly speaking) program which captured the logic (IF THEN ELSE) the neural net is using (defacto using, NN don't have IF THEN ELSE logic except those implicitly embedded in their activation rules) to solve the problem.
That means the algorithm the NN has arrived at is not open to analytical inspection and confirmation, except very indirectly.
This is OK for wide variety of predictive tasks in which human life does not hang in the balance. In medicine, the diagnostic results from NN and even Good Old Fashioned AI expert systems are reality--checked by human doctors.
Neural nets ALWAYS run the risk of coming to the right conclusion for the wrong reason enough of the time to fool humans into thinking it "understands" the problem domain in a way that is analogous to a human. A NN so trained will fool or lull human observers into a false sense of security until that BIG ACCIDENT happens then a post mortum reveals the shocking truth about what the NN was focusing in on to make it decisions.
The Big Idea behind NN is that, through a combination of evolutionary forces and billions of iterations the NN will learn using the same Hebbian activation princples the brain appears (now) to use and that with enough training, the exceptional cases that I am describing will be found and rooted out.
But even in nature, this doesn't happen reliably. Take for example the Australian Jewel Beetle. Over perhaps millions of years, it has of course evolved a robust way to recognize desirable mates and procreate. That is as basic an evolutionary task as you can imagine- it has to work or the species is doomed.
However, the male's algorithm for mating is not as robust as you might imagine. It seems that what males rely on to select a mate is a very, very limited set of perceptual cues. As it turns out, it is looking for big glossy brown curved things. When it sights one, it alights and starts humping away.
Well, Austrailian beer bottles fit this description *and fit it better than the female of the species*. People toss empty beer bottles in the outback and the result is the male beetles prefer the beer bottles to such a degree that the beetles were going to go extinct. Austrailia had to pass a law to change the appearance of its beer bottles.
http://blogs.scientificamerica...
This is a cautionary tale to those who think evolutionary forces produce only *robust* algorithms. What evolution actually produces is *good enough so far* algorithms. What well trained NN produce are similarly good enough algorithms. In both cases we have to do science to try to get at what it is they are relying on- what features they are *really* trained on. And we don't know there's a problem until tragedy happens and we don't know how ridiculous the problem is until we do science.
This is different from procedural programming which, the Halting Problem notwithstanding, CAN be analytically examined for correctness. Procedural type programming plus sensors is what runs water stations, trains, planes etc. The military does use NN to try to recognize things but it has humans making the final decision and when the missle gets launched, it's not left to a NN to decide where to finally land.
Moreover, self driving cars under the control of a NN can and will be attacked by the usual miz of 14 y/o kids, pranksters, criminals and terrorists