You have to keep everyone’s inherent biases in mind when evaluating what they say.
Sensor fusion theory says the more sensors, the better the accuracy (it should not get worse if done right).
Why is Musk so against LIDAR? Well for one it adds costs and looks ugly. And not currently available to practically put on consumer vehicles in mass production.
How is he supposed to sell cars today with the promise of self-driving tech if they don’t even have sufficient sensors? I’m not saying that is actually the case, but you can see he will be biased against needing LIDAR. (Even if years down the road Tesla adds LIDAR, he won’t say that now)
On the other end, basically every other company (Waymo, Mobileye) cannot accrue the amount of data and edge cases that Tesla can. In data science (especially difficult problems), you want lots of data, especially useful data.
These companies cannot get data that way, so they will be biased against approaches that claim you need lots of data to get working. Tesla of course will go down this route because they can.
Shashua saying it’s brute force…absolutely! Guess what is a great way to get edge case data? Build a model that’s 70% accurate. Deploy it. Have it make predictions. Only when the predictions are wrong, collect the data. Retrain model to get 90% accuracy. Deploy this model. Rinse and repeat.
Each cycle, you will only get more and more specific edge case data that the model did not do well with.
Imagine you are at 99.999% accuracy, but you need to get to 99.9999999% accuracy. How are you going to generate enough useful data to get there? You need a lot of cars driving around collecting data because the chance they get something useful is very sparse.
This is a correct data science / algorithm approach.
Guess which is the only company that can do it…
Whether LIDAR or other sensors are needed I don’t know. But Tesla’s approach for their given sensor suite is correct.
)