"Safe AI", i.e., learning with machine beats machine learning alone?

Recent headline news, 1) “Uber halts self-driving tests after death in Arizona” and 2) “Florida bridge collapse” are unfortunately one accident too many. First, our thoughts shall go to the families of the victims.

Next is call for action on continuous learning within the context of AI and machine learning. Both events prompt us to take a closer look at how we learn from data. Over the coming months, please feel free to share your ideas (from geoscience perspective) in this blog. As a starting point, consider the following:

Q1 if neural network (NN) algorithms were used in the Uber self-driving test image recognition, what is the resolution (number of pixels) at which training and testing were performed?

(Concern - NN may come up with different answers on features using input of 2-ms vs 4ms seismic data. In contrast, most of us can enjoy a movie and follow the plot at 720p just as well as 4K ultra-HD TV. How about those 2Mb pictures we used to share on feature phones? We won’t confuse a dog with a person. For fun and motivation, https://www.wired.com/sto...oving-tough-to-fix/)


Q2: if cracks were identified days before the FIU bridge collapse, could that be telltale sign of massive microcracks already coalescing around the tips and integrity of materials around those cracks were seriously compromised?

For serious read, https://www.sciencedirect...pii/0020768388900315
Microcrack coalescence and macroscopic crack growth initiation in brittle solids

Food for thought - whether we are after deep or shallow learning (just marketing label based on the number of NN layers), forging a critical link between machine learning and first principle is the key to unlock AI potential safely.