Tesla will hold a self-driving investor day event on the 22nd of this month, and hopes that it can provide some details of its own research and development work on self-driving at this event;
The realization of fully autonomous driving has enormous financial implications, making certain technical issues all the more important;
For investors, the key question now is whether Tesla can take advantage of the massive amount of data it trains on in the real physical world to achieve better self-driving than its rivals.
Recently, Tesla announced that it will hold an autonomous driving investor day on the 22nd of this month. The following is the content of the announcement:
“Tesla will have a series of exciting developments in the coming weeks and months, and we can’t wait to share them with you. Tesla will hold an investor day at its headquarters in Palo Alto, when the , we will help you gain an in-depth understanding of the company’s autonomous driving technology and roadmap.
We are offering investors a test drive where investors can experience our latest self-driving software first-hand, including some features and capabilities that are currently under active development. At the same time, Elon Musk, Stuart Bowers, vice president of engineering, Pete Bannon, vice president of hardware engineering, and Andrej Karpathy, senior director of artificial intelligence, will all be present and speak. ”
Here are five questions I want investors to ask Tesla at the Autopilot Investor Day:
1. It is clear that Tesla is using deep supervised learning for computer vision tasks. But for the actual driving task – route planning and driving policy (to use Tesla’s own terminology), is Tesla using imitation learning, reinforcement learning, or both? In imitation learning, the neural network learns to drive by observing human driving behavior and correlating the perceived information with the driver’s behavior, and in reinforcement learning, it learns by trial and error (usually by running simulations).
2. No matter which solution Tesla uses for driving tasks (as opposed to computer vision tasks), when companies like Google Waymo are still struggling, why does Tesla think the solution it chooses effective? Why does Tesla have its own unique advantages in implementing this scheme? For example, if Tesla chooses the simulation learning solution, then it has a unique advantage because it sells a relatively large number of cars, then we hope that Tesla can provide us with a proof why the imitation Learning will work. And why doing pure reinforcement learning in simulation alone doesn’t work.
3. How far is Tesla at solving the necessary computer vision tasks? How does Tesla set a standard to measure the completion of this task? Is the rest of the work just adding more labeled data to the training set?
4. How accurate is Tesla’s sensor suite at acquiring depth information about surrounding objects? LiDAR is sometimes touted as an advantage in depth mapping, so what are the strengths and weaknesses of Tesla’s sensor suite compared to LiDAR?
5. Does Tesla see potential in end-to-end learning? Or is it still far from the practical stage of end-to-end learning? How to combine end-to-end learning and unsupervised representation learning? In end-to-end learning, a neural network is trained by imitation learning or reinforcement learning to generate actions based on raw sensor input, without the need for human labels at any stage of the process. In unsupervised learning, a neural network learns from past data to predict future data without any human labels.
Right now, we know almost nothing about Tesla’s technology strategy. We can certainly speculate on its technology strategy based on some of the evidence we have now, but we’re not sure exactly what Tesla is doing behind the scenes. It’s possible that Tesla is just hosting an ordinary self-driving investor day this time, and it is very secretive about the technical issues I listed above, but I hope that it can use this event to be open and honest with investors. Explain in detail what Tesla is doing now, and talk about why the technology strategy it is executing is the right one.
According to the financial model released by ARKInvest, if Tesla can launch a fully self-driving car in the next five years, its stock price will increase by about 8 to 14 times on the current basis. Although different analyst firms have different financial forecast models for Tesla, many companies believe that the potential long-term value of autonomous driving greatly exceeds Tesla’s current market value. If a company can successfully deploy fully autonomous driving technology, then investing in it blindfolded is a simple and crude logic. But the question now is, from the perspective of technical feasibility, whether fully autonomous driving can be realized, and if so, which company will realize it. That’s why it’s important for investors to understand Tesla’s self-driving technology strategy.
Now, the key question facing Tesla investors is how to draw a conclusion from the answers to the above five questions, that is, whether Tesla can really use the huge training data of the 450,000 cars it has sold to achieve the comparison. Other companies have better autonomous driving performance. What sets Tesla apart is its vast training data set. But such resources can determine Tesla’s competitive advantage only if the winning technical solution requires a large amount of training data that does not require the bottleneck of expensive and slow human labeling. Imitation learning and end-to-end learning are such technical solutions. The opposite approach uses hand-coded rules to tell the self-driving car what to do, without using any training data from actual driving. An alternative is pure reinforcement learning, which uses computers to generate training data in a simulated world, rather than using data from the real world.
Therefore, whether Tesla has a strong competitiveness depends on whether its huge training data from the real world is the decisive resource for the final winning plan. Whether Tesla’s real-world data is a decisive resource depends on whether the winning solution is data-intensive (like imitation learning or end-to-end learning) or data-insensitive (like hand coding or Pure reinforcement learning in a simulated world).
In other words, Tesla’s competitiveness actually depends on two aspects. First, what is the technical route Tesla is currently implementing? Second, whether the technical solution it chooses is correct. Then, the corresponding information that investors need to get from Tesla is, first, an explanation of the current technical route, and second, a demonstration of the correctness of the chosen technical solution.
Based on a series of investigative reports, I believe Tesla is using a “simulation learning approach” to advance the self-driving mission. I am optimistic about this solution, mainly for two reasons. First, DeepMind’s AlphaStar project compellingly demonstrated the power of pure imitation learning and the use of imitation learning to augment reinforcement learning. Second, experts from Waymo and UberATG have also publicly emphasized imitation learning as a very promising solution to autonomous driving tasks.
Also, I’m skeptical about hand-coding schemes because, so far, it’s had little success because humans
Sometimes it’s hard to formalize complex tasks into an established set of rules, so I’m on the fence. Also, I’m skeptical about using pure reinforcement learning in simulations, because in my opinion, it’s necessary to have a realistic model of human driving behavior in order for the neural network to learn how to interact with other drivers in complex situations. interactive. And to create a real human driving behavior model is equivalent to creating a fully automatic driving car, so there is a question of chicken or egg.
If we can get enough information at Tesla’s self-driving investor day to demonstrate the advantages of imitation learning schemes, and Tesla will demonstrate why its scheme works and others do not, then the event will be Is a very important event, personally, I think it is more important than Tesla announced its latest quarter earnings or how many vehicles it delivered.