4 ms·
If you watch the presentation at the Tesla Autonomy Day, the dude in charge of their deep neural network explained that they think basing the whole system upon
by olau 7y ago
If you watch the presentation at the Tesla Autonomy Day, the dude in charge of their deep neural network explained that they think basing the whole system upon cameras is necessary since HD maps and LIDAR are too limiting. Basically, the network needs to be able to see what people are doing to drive safely.
They've devised a method of upon training data for their network and is basically iterating through the problems, finding training data and counter-training data, as far as I understood the presentation.
Whether this is going to work or not, we shall see.
- pas 7y agoHave they solved the world model problem, or mentioned anything about that? If not, they are lost in noise. Video streams are so full of maybe signals, and ML loves that. As long as they don't have a fundamental (sub)system that knows how much space it needs to safely proceed at the current speed, how the 99.99% of the roads of the world work geometrically, and how to infer this from visual data, they are in trouble. (That said, I guess Karpathy is perfectly aware of this.)
- jaimex2 7y agoI highly recommend watching the presentation. It's quite a deep dive into their entire plan including solving the world model problem.