I feel like this section is rather disingenuous for the article author to just drop without mentioning that this is how all machine learning models are trained. The idea is that now (and for the next year or whatever) it’s trained manually until the system is good enough to do it on its own with a good enough accuracy rating to not lose money.
Now, since Amazon is shuttering this, it’s totally possible that they determined they’d need too many years of training data to break even, but at the very least this is standard industry practice for any machine learning model.
I thought so, top, at least initially. However, it sounds like the out sourced cashiers essentially tallied up the orders after the customer left and sent the bill. How much correction had to be applied to the automation remains to be seen. The biggest issue with systems like this is tracking between scenes and angles. It could be the humans in the loop were there to resync the metadata of each tracked object (i.e., customer) as they moved isle to isle or they removed a jacket or whatever.
I fully agree that the author is being super disingenuous here. However, I don’t think Amazon is fully shuttering the program because they’ve stated they’re switching it over to a cart-based system which is already been proven to be successful in the trade and doesn’t include the high ceiling requirements, their biggest issue is adoption from other retailers, and switching over to that system will lower the ceiling(no pun intended) for entry
It’s possible the article is referring to how they got the training data, but as we’ve been learning from recent reporting, places like cruise were actually hiring people to monitor the “autonomous” systems.