Maybe they're only sending small updates for the road you're driving on rather than large updates that include all the road corrections for an entire region or the country. That might explain why you don't see a big uptick in data while parked on wifi. It probably only takes a few kbs of data for the road you're on for some significant number of miles. I suppose the counter to that proposal is that it doesn't cost Tesla anything to use your wifi so why not upload a huge update overnight while on wifi?
You could be right. Obviously I don't have insight into the structure of their model. Initially I figured it would be like most problems, in that you'd make a model, train it on the massive data, and intend to have it generalize well to new conditions. It's definitely possible it's region specific and in a double layer, as postulated above. Having a geolocation-fractured model might make sense because the model wouldn't have to generalize as greatly and could better fit the conditions in the region.
Getting too specific (the idea of the road ahead for some miles) has the sound of hand coding to it, and I think we all agree this is some kind of machine learning algorithm. So I'm not too sure about that piece. I would hate to think that there are engineers hand coding specific highway exits as failsafes to a deep learning network.
To take the right-hand exit example, I'll give a very simplified and high level idea of how a machine learning system could work. The actions are based on
features which have
weights. There are many features but let's really pare it down for the purposes of this example.
Features:
- Left line location
- Right line location
- Center of vehicle ahead
- GPS longitude
- GPS latitude
Each of these features is given some weight and contributes some influence on the behavior of the model at any given time. Perhaps initially, for a given set of features, the weight of the right line location was higher than the left line location except in cases where the center of vehicle ahead was populated (existed). Then, AP might have a tendency to lean more towards following the right line. However, imagine that every time it does that at a right exit, the driver disengages it and keeps the car on the highway. Then, given some GPS coordinates and/or other features, the weight for the right line would decrease in the updated model. Every occurrence of a driver keeping the car on the highway would continue to lower the weight, until it reached some optimal value.
The chart for the right line location weight might get very high in locations where there is no exit, and then rapidly decrease every time you pass an exit. This is the amazingly cool thing about these predictive models - how they learn by creating a "cost" to doing a certain thing, and minimizing that cost.
Leaving GPS out of it, it's also possible that a model could just figure out that when the right line curves away when the left is going straight, it is probably an exit and could lower the weight. The upside to this is that it's generalizable to other locations that haven't been seen by the model.
The key piece of this is that these models require processing power and time to build. So I do not think they are continuously updated, unless they are treated as a kind of time-series problem, and I can't see how that would make too much sense.
Sorry for the long post. TL;DR - computers are cool.