That thesis fell apart from the start because it is based almost entirely on conjecture, cherry-picking, and wishful thinking.
I mean this is just a general assertion. Not something that can be specifically addressed with evidence or reasoning. ¯\_(ツ)_/¯
I like it better when people make specific claims I can respond to.
But Occam's Razor says that you should really look at the reliable information we have -- i.e., past and present performance of the vehicles themselves
If we were to take this to its logical conclusion, wouldn't we conclude that fully autonomous vehicles are impossible, since they don't exist today?
Talking about the future of technology requires conjecture about things that exist today because future technology doesn't exist today. But that doesn't mean it's just free-wheeling science fiction. We can look at an existing technique like imitation learning, for instance, and wonder what might happen if it were applied in a new domain or at a new scale.
From just a couple of posts up you sound very sceptical of every other autonomous manufacturer — ones with far better known merits than Tesla in this sphere — and then you do this with crumbs of theoretically positive Tesla news (for your thesis, that is)... go completely overboard in my books with the excitement and importance...
Let's be clear about what we're talking about here. Is Mobileye, or BMW, or any other company (besides Tesla) collecting state-action pairs from production cars? If so, I would like to know! If you're aware that they are, please provide a source. If you can't provide a source, then how do you know it's happening? Or do you even claim it's happening? This is a very specific factual question, not a general assessment of progress or capability.
Conversely, Bladerskb previously asserted that there is no evidence of Tesla collecting any state-action pair data from the customer fleet, and verygreen (as I understand it) stated that, actually, Tesla has set up collection of the NN mid-level representation (state) and driver input (action). Again, this is a specific factual claim.
AlphaStar was trained on 500,000 games
I may be wrong, but AFAIK, the dataset is much larger. We will know for sure when DeepMind releases its paper. But for now, there's
this source:
"Blizzard will make about 500,000 more games available each month."
Secondly we don't know the amount of end to end imitation learning data needed for driving. We do know that AlphaStar wasn't trained on billions of miles of equivalent continuous driving data. ... Case in point, you don't need 'billions of miles'.
AlphaStar and autonomous driving are just an analogy. StarCraft and driving are different tasks. I think you were right the first time when you said we don't know how much data might be needed.
In my mind, the point of the analogy is just that AlphaStar shows imitation learning can handle complex, long-term, real time tasks in a 3D environment with elements of strategy, tactics, and multi-agent interaction. I don't think comparing hours of StarCraft play to hours of driving allows us to predict exactly how much data is needed for driving. It's just an analogy. For example, AFAIK, StarCraft doesn't have the long tail that driving does.
So no Tesla siphoning ~0.1% of data isn't a AlphaStar approach and is nothing like AlphaStar at all.
As rnortman put it, the main point is that Tesla has:
a resource they can tap for training FSD when they're ready to do that
I feel like we've already been over this. Post-HW3 imitation learning is going to be much more interesting than pre-HW3 imitation learning. Not necessarily
immediately post-HW3, but HW3 will enable improvements in the perception NNs, which we agree is required for effective imitation learning.
If Tesla is able to solve reliable vision at some point they could collect more but solving that seems a long way off and continues to be hampered by a more limited sensor selection than other autonomous players have so even this route is not there anytime soon.
To
quote Mobileye:
"While other sensors such as radar and LiDAR may provide redundancy for object detection – the camera is the only real-time sensor for driving path geometry and other static scene semantics (such as traffic signs, on-road markings, etc.)."
The sensor modality Tesla does not have — lidar — doesn't help with things like signs, lane lines, stop lines, cross walks, the colour of traffic lights, brake lights, and turn signals.
Look, the whole thing that made AlphaGo and AlphaStar (the
@strangecosmos thesis) such a hero was the continuous feedback loop of the system playing itself time and again.
You're referring to
reinforcement learning via
self-play. But I have already addressed this point:
- Per DeepMind's estimate, AlphaStar attained roughly median human performance on StarCraft using imitation learning alone.
- Researchers such as two at Waymo have pointed to imitation learning as a way to create "smart agents" that could enable reinforcement learning in simulation for autonomous driving.
I'm not saying that I can predict the future and that applying the same techniques to driving as to StarCraft will certainly work. I'm just saying that it's an intriguing idea — it seems promising enough to try, and I hope Tesla does indeed try it and that we can see what happens.
Particularly if the alternative is the 15-year-old approach of hand coding cars to drive, we need to explore new frontiers in machine learning for robotics if we're going to overcome the hurdles to fully autonomous driving.