There are going to be new neural networks. Andrej Karpathy confirmed this on the
Q3 2018 earnings call:
“...my team trains all of the neural networks that analyze the images streaming in from all the cameras for Autopilot. For example, these neural networks identify cars, lane lines, traffic signs, and so on. The team is incredibly excited about the upcoming upgrade for the Autopilot computer which Pete [Bannon] briefly talked about.
This upgrade allows us to not just run the current neural networks faster, but more importantly, it will allow us to deploy much larger, computationally more expensive networks to the fleet. The reason this is important is that, it is a common finding in the industry and that we see this as well, is that as you make the networks bigger by adding more neurons, the accuracy of all their predictions increases with the added capacity.
So in other words, we are currently at a place where we trained large neural networks that work very well, but we are not able to deploy them to the fleet due to computational constraints. So, all of this will change with the next iteration of the hardware. And it's a massive step improvement in the compute capability. And the team is incredibly excited to get these networks out there.”
I'm wondering if part of their plan is to wait on announcing the new hardware so that they have time to gather miles and train the new NN with fleet data. It's not great if they announce new hardware that is supposed to be amazing and then say "oh, well actually it's not going to be that good for a few months". Reminds me of HW2...
Karpathy’s quote above suggests that all the training is already done. After all, the sensor data captured from cars is uploaded to Tesla’s servers, so it doesn’t matter what computing hardware the cars have. The Hardware 3 NNs can be trained on the exact same dataset as the Hardware 2 NNs. If there is a delay before Autopilot is active on Hardware 3 cars, it will probably because Tesla is just testing that the new hardware and NNs are safe, and not anything to do with training.
At least not with regard to perception. Path planning/driving policy might be a different story. If Tesla is switching over from a hand-coded approach to a neural network approach, and needs lots of human driving data for imitation learning, but needs Hardware 3’s perception networks to get that data with high accuracy (not the raw sensor data but the metadata or mid-level representations), then possibly that could cause a wait before Autopilot is active. But I don’t know why they wouldn’t just ship the old hand-coded software until the new neural network is ready.