Agree with your point regarding computing-at-the-node, but think those will be edge cases due to…

Agree with your point regarding computing-at-the-node, but think those will be edge cases due to regulation, privacy, and bandwidth (in that order). For privacy, see Apple’s photo efforts in iOS 10. For regulation, think HIPPA or cross-border issues.

Privacy and regulation overlap when it comes to France and Germany’s pushing ‘right to be forgotten.’ If I choose to void all my data Google has, they’ll have to rebuild all models which learned from my very, very small contributions. Due to the models being black boxes and the difficulty predicting who will opt out, I imagine they might take a hybrid approach. Super interesting thought experiment, though.

For bandwidth though, I think these cases will revert back to the cloud if the bandwidth problems are solved. The value of the data improving models for all (I certainly want my car to learn from your car’s experiences) is too great. I imagine in most cases where bandwidth is an issue, some processing will take place at the node to reduce the data significantly before it is shuttled back to base.

But then again, you certainly know much more about the self driving case than I!