[jessaustin]

Aren't there fully self-driving vehicles undergoing testing on normal city streets with stoplights? We could say that's a different mode than high-speed highway driving, but there are a number of locations where those modes blend into each other. Are the experimenters extra careful to avoid those locations? How long before the tech exists to address this issue?

[gambiting]

Sure there are, and yes, they are different sets of problems though. Have a look at the British Tesla Driver Youtube channel, some of his videos are eye opening. Basically the car is in full autopilot mode, approaches an intersection, correctly slows down, waits for its turn, starts moving.....and in the middle of the turn goes BEEP BEEP BEEP and disengages entirely because the road markings aren't there and it wasn't entirely sure where to go. And now of course you're in a moving vehicle that's heading for a collision with someone else and requires IMMEDIATE attention to continue. One could(and I'm sure will) argue that the system "shouldn't be used this way". But that's a moot point, if the system is there and lets you do this, then people will use it this way.

"How long before the tech exists to address this issue?" I'm not sure if that's a problem with tech as such. We have fantastic cameras, yet famously Google's best image recognition algorithm just couple years ago would reply, with 100% confidence, that a sofa in a zebra print is in fact a Zebra, after all the stripes are there, it has 4 legs.....it must be a zebra.

So in my(personal) opinion, self driving will face the same challanges image recognition has faced - we will rapidly get 90% of it right, then the last 10% will be a massive pain to get right for decades if ever.

[Isinlor]

It's an active research field. E.g. from October 2020: Calibrating Deep Neural Networks using Focal Loss: https://arxiv.org/abs/2002.09437

> Miscalibration - a mismatch between a model's confidence and its correctness - of Deep Neural Networks (DNNs) makes their predictions hard to rely on. Ideally, we want networks to be accurate, calibrated and confident. We show that, as opposed to the standard cross-entropy loss, focal loss [Lin et. al., 2017] allows us to learn models that are already very well calibrated. When combined with temperature scaling, whilst preserving accuracy, it yields state-of-the-art calibrated models. We provide a thorough analysis of the factors causing miscalibration, and use the insights we glean from this to justify the empirically excellent performance of focal loss. To facilitate the use of focal loss in practice, we also provide a principled approach to automatically select the hyperparameter involved in the loss function. We perform extensive experiments on a variety of computer vision and NLP datasets, and with a wide variety of network architectures, and show that our approach achieves state-of-the-art calibration without compromising on accuracy in almost all cases.

Calibration will be practically solved in couple of years. Then a bit longer for addressing adversarial robustness.