Reductionist
Member
I don't think it does? The performance of any given NN is based both on architecture and the training data used. Tesla currently has such a high performance system not because their architecture is all that special (it isn't, see talks by Andrei Karpathy for details), but because they have access to the best/most varied training data for the task at hand *by far*.FWIW if teslas vision system is as good as that post makes it sound, then it has multi-billion dollar potential in probably a dozen different industries (im being very conservative there), that have nothing to do with cars or energy.
Its so frustrating as a brit that I cant get a model Y yet (I still have AP1 on my 2015 S), and that the EU stupidly dumbs down autopilot here, presumably waiting for germans to catch up...
If they switch to a truly innovative NN architecture in the future, this in of itself won't have much value in other vision applications beyond automotive, unless you can also figure out how to get a training set that's similar in quality to the one Tesla get's through the fleet for their application. This seems very difficult to me for many of the applications that CV would be useful for, so architecture really is only half the battle, if even that.