Hacker News new | ask | show | jobs
by oh_teh_meows 3776 days ago
Not in the traditional individual sense, I think. All AI drivers across the whole world will seamlessly exchange and synchronize with each other their learning experiences, resulting in self-driving vehicles that improve at an exponential rate.

It's basically an AI collective, whose 'brain' scatters across data centers around the world.

5 comments

No way will the driving get exponentially better with the dataset size, probably closer to logarithmically actually, really fast in the beginning and progressively much much much slower
To be honest, I'd be worried about security vulnerabilities from a system that was exchanging and synchronizing data in the cloud. It seems unlikely that the first generation of self driving vehicles will be able to update its behavior in real time from other vehicles' experience. If there is any cloud connectivity at all, I'd expect it at first to be solely outgoing data that is sent to and aggregated by some central brain. Then, when the owner of a vehicle wants to update its AI, they manually download a certified patch or bundle of data that incorporates what the central brain has learned from all the cars. These upgrade patches would probably have to be constantly vouched for by debug vehicles that drive all over the streets of the world before being released.
>Not in the traditional individual sense, I think. All AI drivers across the whole world will seamlessly exchange and synchronize with each other their learning experiences, resulting in self-driving vehicles that improve at an exponential rate.

That's a nice picture of free-form learning from inputs, but the reality could be quite more rigid.

I'd expect something more like gathering "black box" style information of the situation (the analogous of a stack trace or core dump), and having programmers trying to improve the algorithms, while also taking into account all the safety issues that could go wrong, than some "let's change my behavior because I was penalized" neural-net/ML way.

It's already happening for google's driverless car program. I highly recommend the Ted talk about how driverless cars see the world. Fascinating.
Wouldn't left side driving in other countries screw the learning?