[sheepshear]

If the list is of objects they can't yet reliably detect, then how would they implement your suggestion to detect those objects? I'm sure they have a lower tolerance for false negatives, so of course the known problems must be hard-coded.

[londons_explore]

Lets imagine there is a stop sign that is half hidden by a bush and pretty faded, so the stop sign detection logic says "only 5% chance it's a stop sign". That in turn isn't enough to make the car stop.

The hardcoding approach says "This is a sign. 100% sure.".

The vector approach says "There's a hard to see stop sign around here, boost up the probability of anything stop-sign-ish a bunch".

The difference functionally is that nothing in the real world is ever 100% certain. So you should never tell any bayesian machine (which a neural network effectively is) that anything is 100% true.

The vector approach I outlined is far more general than the above though - it allows any behaviour of the car to be tweaked automatically or manually. location-specific vectors can be learned from data, and/or put in by operatives. The way the neural net trains, the meaning of a vector could 'evolve' too - for example, whenever a human puts in that there is a hidden stop sign, the neural net might learn that that means other human drivers might occasionally fail to see the sign and stop in those locations. Even though it had never witnessed a human failing to stop in this specific location, it has learnt that is part of the meaning of this vector.

[sheepshear]

I'm saying it's very likely that the business has exactly zero tolerance for the detector missing those signs while there are known, unresolved issues. They can't have something happen when they "knew it could happen".

After "resolving" the issue, then they can resume tolerating the normal probabilities.

In the US, it might be more of a precaution. In Europe, it could be more of a legal obligation. Either way, I doubt engineering has a say in the matter.