Hacker News Reader

Comment by Fricken | Hacker News Reader

One can scale up the first 99% of an autonomous driving os pretty fast just using pattern recognition, but the devil is in the edge cases.

We haven't heard much from Google because they've been programming and validating all the unglamorous little problems that training data and a neural net can't solve alone. 4 way stops, passing motorcyclists and stuff like that, where old fashioned hand coded features are necessary to adress situations in which specific kinds of reasoning is necessary.

Tesla has to solve all this boring stuff as well before they can offer comparable capabilities in an AV.

View on HN

>old fashioned hand coded features are necessary

I don't believe your hand coded features will outperform ML trained over a billion miles, even in the long tail rare events.

Sure ML is bad at rare events, but hand coded features also map poorly to such rare events, because it's unlikely you'll be able to imagine them all, or even manage to handle categories of rare events well.

Furthermore, I believe it would be possible to train an ML based rare-event detector, which simply detects the difference between "regular driving I'm familiar with" and "something funny is going on, I haven't seen this much in the training dataset", and then refer to a remote human to resolve the issue.

> and then refer to a remote human to resolve the issue

Successful remote control is a whole different project, very different from self-driving. In order to give the remote pilot enough situational awareness, you need plenty of streamed imagery (recreated or recorded). And you need a feedback loop with a very small latency. Not saying it couldn't be done, but it won't happen "organically" by cars becoming self-driving.

That's the benefit of having 100,000 cars with all the sensors but humans performing all of the edge cases. Tesla can take data from all the cases where there was a deviation between what the car calculated it should do and what the driver did.

I don't think hand coded features are necessary. No one hand codes humans to drive.

Reply on news.ycombinator.com