I like Filip Piekniewski's take on this [0] - deep learning is appropriate for environments where the cost of mistakes is low (or the errors are reversible) and you're able to 'average them out'. Self-driving cars are absolutely not in that category. "If your errors are irreversible and you can't average them out, and your application is in open domain where there will never be a shortage of out of distribution sampl…
> For me it's obvious the real issue are the edge cases - of course a self-driving car does fine in 99%+ of cases, but those small % of outlier cases are the ones where people potentially die. I suspect an entirely different approach is required to supplement/supplant the deep learning one and fully autonomous vehicles are still quite some way off. Why is that every time the safety of self driving cars is brought up,…
Also it is not at all obvious that self driving cars will do better than humans. Perhaps they will, but it isn't just about numbers. As I alluded to in another comment, imagine if they had 50% of the human death rate but it was all small children and occurred seemingly randomly. Do you think that would be acceptable and a better outcome? These things are less black and white than they might seem.
Another factor is that with deep learning it is fully a black box. It is very difficult to understand WHY a particular course of action was taken, whereas in human cases you can usually determine this. The why might also be limited, e.g. 'human child was mislabelled as stop sign' or similar, which just causes further questions.
The fact that you have scenarios the car will definitely not know how to handle combined with a deep level of inscrutability makes it difficult to have confidence in issues being resolved or the technology improving on a safety basis.