Self Driving Cars Still Don’t Know How to See | Hacker News Reader

Self Driving Cars Still Don’t Know How to See

theatlantic.com

14 pointswskinner8 comments

The article starts off reasonable but then seems to veer off into unsupportable assertions about human reaction times when presented with danger and how self driving cars will kill more people.

And some of the comments about processing power seem at odds with Waymo videos showing it tracking all objects in the scene and making predictions, even non-moving ones, such as emergency vehicles on the side of the road.

This shouldn't be shocking to anyone. Computers still can't consistently identify a banana when it's the only thing in the picture, despite the fact that a small child could with perfect accuracy.

While the human brain is rather slow on response time, it's pattern recognition is still far superior to even the best algorithms we have today.

It's actually been a few years since computers can outperform humans in image classification. Look into the imagenet competitions since 2015 for more details.

This may be true with large datasets, but with small datasets it's no contest. I don't have to show a kid a thousand slightly diffferent bananas in different environments to teach them what a banana is.

To be fair, on the one hand we are training computers on thousands of static photos of bananas.

Kids are trained with continuous live action of bananas they are turning around, seeing with binocular vision, seeing from all angles, pulling apart, etc.

Kids are taught with massive visual data.

Maybe not a thousand, but the kid also has a wealth of understanding of greater concepts which the banana fits into even if never seen before. When there is such fundamentally different architecture of learning its kind of absurd to compare, at least to take such a comparison seriously. When AI can build a full world model then we'll be somewhere, I think reasoning is a dependency of that. We don't require thousands of examples of an object to be able to recognize it partially because we aren't learning from scratch. You're able to look at a banana for the first time ever and automatically fit it into many already known classes (object, edible, yellow, elongated shape) based on reasoning about it vs past experience, so the new information learned to make it permanently recognizable is very small.

I have looked into it and I stand behind my statement.

It seems like the article is ignoring the countless pedestrians that have been killed by human drivers. Obviously self driving cars will make some mistakes, but I believe (correct me if I'm wrong) that self driving cars already have a much better safety record than human drivers even with the as yet unsolved problems.

Reply on news.ycombinator.com