I don't think neural networks are wired to "remember" things. In theory, they could be hooked up that way. But your typical convolutional neural network is looking at things frame-by-frame.
In theory, ANNs could have an output layer that passes data from one frame to another frame to assist things. But there's no real programming to "hardcode" something like object permanence into an ANN. You pretty much throw a bunch of data into the system and hope for the best.
NNs are just the first step in the pipeline. Their outputs (detected objects, segmentation, etc) will be piped into other software that builds higher level models.
Considering the path-planning requirements I would be absolutely shocked if Autopilot wasn't build history models and estimated paths for objects around the vehicle (other cars etc).
Agreed; I imagine they use neural networks to detect and classify objects which are then saved into a scene-graph for use in pathing.
I expect what happened was that they trained their NNs for improved detection in one area but unknowingly reduced it in another. Perhaps now it can detect tricycles 99% but road barriers went down to only 30%. Having worked with NNs it's very common to see gains in one domain which come at a cost of reduced performance in another.
They must have known. I haven’t worked with NNs for a few years but I don’t believe the methodology has changed where you would stop testing it over different sets of training data.
Barriers are a pretty big part of driving on roads and highways and the only reason it would have been unknowingly reduced would be if they just weren’t testing the NN against data with them.
Thanks (to all in this subthread). I've skimmed through this doc, I have watched https://vimeo.com/274274744 linked from original reddit discussion and I have lost all remaining faith in AP as it is today. This should be in a closed alpha version, not anywhere near paying customers and marketed as FSD ready/feature ready as it's none of that and won't be for many years. I expect there will be at least a generation of Tesla cars sold with FSD-readiness package that will never see FSD in their lifetime.
In theory, ANNs could have an output layer that passes data from one frame to another frame to assist things. But there's no real programming to "hardcode" something like object permanence into an ANN. You pretty much throw a bunch of data into the system and hope for the best.