Worth watching the Tesla Autonomy day video to see how some of this works
Skip to about 2hrs 18mins for some of the discussion of deriving depth data from vision
I agree that material should explain something, but imo when you cut to the chase, all he says is 'it just happens in the NN'. I suspect that NN's have a closer link with monkeys and typewriters than FSD owners and Tesla would want to own up to!
A while ago I found a fun demo / simulation of a NN being trained to do basic number recognition. While the NN was being trained you saw it starting to gain confidence in its recognition. When trained, it was 100% correct for all numbers, but the model showed you the %age possibility that a given written number could be each possible digit 0-9.
You could then adjust the given 'number' adding dots, smudges or deleting parts and see how the NN's predictions changed. So a small smudge near the centre of a 0 caused an increase the the possibility that the 0 was supposed to be / could be an 8. There were many predictable outcomes for other extra marks near the subject character.
What suprised me was that removing a seemingly insignificant part of the character would reduce confidence for all possibilities. Ie: the presence of the removed line was significant in confidence for all numbers.
Of course this model was super basic, but the problem I see with NN's is that it must be near impossible to know what attributes they are building up reliance upon to differentiate between objects etc.
Nature is really good at being random in truly unfathomable ways!
Then you have sensors that have unavoidable bias / perspective that could also build in unexpected dependencies / associations in the NN.
From playing with simple autonomous robots, I recognise the dilemma of more sensors giving more data but also giving greater chance of erroneous data. One infallible, universal sensor would be so much better than less consistent ones having to argue about who's right! Adding over-riding safety loops that do something in the case of a suspected system error / un-handled condition seem like a good safety feature until you realise that their predictable behaviour often comes from simplicity and a simplistic view is often not the best one!
From a geek view point it is amazing what Tesla has already done and I would say nothing against that. I did not select my MS because it has FSD, but it met all my other parameters including price so I ended up with FSD and enjoy having a small participation in the development cycle.
Regardless of the reasons behind changes in use of Radar and how that might relate to phantom braking, if Tesla Vision can get it right all the time, that would be ideal. It does feel that we are still at the very earliest stages of a potentially very long journey.
Last edited: