Only Tesla really know, but I think a small part of what Lattner did was improve the performance (which would be in line with verygreens findings that the inference model is less than half the size, and presumably more efficient - Lattner mentioned significant improvements in performance), but more significantly, he probably oversaw the installation of the 'learning machine' - the complex data harvesting and number crunching processes that work at Tesla HQ to bring these improvements. That's something we'll never see, or really know about, but a massive part of the overall puzzle; and very, very complex to implement well.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Firmware 8.1 - Autopilot HW2
- Thread starter HankLloydRight
- Start date
Only Tesla really know, but I think a small part of what Lattner did was improve the performance (which would be in line with verygreens findings that the inference model is less than half the size, and presumably more efficient - Lattner mentioned significant improvements in performance), but more significantly, he probably oversaw the installation of the 'learning machine' - the complex data harvesting and number crunching processes that work at Tesla HQ to bring these improvements. That's something we'll never see, or really know about, but a massive part of the overall puzzle; and very, very complex to implement well.
3Victoria
Active Member
Being smaller, it may not be quite as good as the previous. But, removing the 'cruft' would leave more potential space for additions to the model, which is a very good thing. I have high hopes with the addition of a very good AI researcher.
Bet TSLA
Active Member
Perhaps it's just being stored in compressed form. This sort of thing becomes smart when you have a faster CPU and space or bandwidth constraints.
That doesn't necessarily follow. A smaller model has less capacity, but it's also less prone to overfitting to the training data(making it work very well on the exact situations it was trained on and poorly on everything else). In general, the model with the absolute best performance will be the smallest one possible that still has sufficient capacity for the task at hand. Finding where that threshold is is both difficult and architecture dependent.
Does not look like it's compressed. I tried gzip on the data and new 12M compressed into 7M, old 29M compressed into 27M.
Structure is a lot more visible in the new one, though as in a lot more text strings).
From what I gather, there's a lot that goes into presentation of the data to the learning network. With the behavior being so different than before in a number of ways, and the model being so much smaller, I wonder if they fundamentally changed the whole thing.
Similar threads
