Hacker News Reader

Comment by SQueeeeeL | Hacker News Reader

This article is behind a subscribe wall, which isn't very conductive to sharing. But upon reading, it seems to be that the technology is just not capable of performing the task. The assumption that deep learning is sufficient to be able to drive a car may be a faulty one.

View on HN

They are doing much more than simply deep learning: https://www.youtube.com/watch?v=j8qypWV2L2g

Deep learning alone has long since been considered a faulty approach (see ALVINN), since it could produce a car that can sort of drive, but nobody knows how, and when a road rule changes you have to retrain it from scratch.

That's not what I got from the two paragraphs it's possible to read. It suggest Uber "is destined to lose the high-stakes race to its rivals, which have demonstrated a lot more headway" which sounds like it is doable, just not by them

Won't it be great if Waymo has to end up licensing the driving tech from Google/Alphabet/WhicheverCoTheyStoleFrom?

Edit: Of course I meant Uber. Doh!

I think you've mixed up some names, given Waymo is an Alphabet company.

Waymo seems to be doing much better than Uber, so it can't be just about what's possible. Waymo are also doing much more than just deep learning. For example a deep learning algorithm might be good for predicting possible paths of tracked pedestrians, but figuring out whether you will intersect any of those paths is better solved by a different non-learning algorithm.

Waymo almost predates the current deep learning fad.

While I'm sure they have deep learning somewhere in the stack, the core of their algorithm appears to be based on more traditional techniques, with kalman filters being the core.

> the core of their algorithm appears to be based on more traditional techniques, with kalman filters being the core.

Interesting! Can you please share a source so that we can learn more?

Reply on news.ycombinator.com