Interesting blog post from the CEO of now defunct AV company.

[Daniel in SD]

The End of Starsky Robotics

 

[diplomat33]

T

Where's the blog post or I am missing the joke?

 

[TydalForce]

There's a blog post there, but it does an after-page-load so it might get caught in your content filters or something.

 

[Daniel in SD]

Where's the blog post or I am missing the joke?

I accidentally posted the thread without the link. I couldn't delete the post so I edited it.
Here the link in case you have caching problems: The End of Starsky Robotics

 

[diplomat33]

I accidentally posted the thread without the link. I couldn't delete the post so I edited it.
Here the link in case you have caching problems: The End of Starsky Robotics

Thank you. I will check it out.

 

[diplomat33]

I accidentally posted the thread without the link. I couldn't delete the post so I edited it.
Here the link in case you have caching problems: The End of Starsky Robotics

This quote caught my eye:

" The biggest, however, is that supervised machine learning doesn’t live up to the hype."

Ouch. This sounds like bad news for Elon's lofty FSD promises. Frankly, the fact that Elon has been promising L5 robotaxis since 2016 and we've yet to get "city NOA" should be a hint that perhaps the camera vision + supervised machine learning approach is not as easy as Elon thinks.

I am not saying that Tesla can't do good stuff with supervised machine learning. They have certainly made a lot of progress with AP since 2016. I am sure Tesla will do more good stuff. So supervised machine learning will do good stuff. I am not bashing AP and I am not saying that Tesla can't do anything. All I am saying is that it will take Tesla longer to achieve L5 autonomy than Elon thinks. Supervised machine learning is just one of many tools. It is not some magic bullet.

 

[J1mbo]

Thanks for posting. Very interesting. I tend to agree that pattern matching alone isn't enough, and that safety is typically only apparent when it goes wrong, hiding the huge engineering effort invested to make sure that doesn't happen.

Reading the blog, I felt that the comments Kaspersky made about getting the architecture right was critical were very relevant, and I imagine that the "hydra" approach they are now taking helps a lot when dealing with edge cases. Slightly OT, but does anyone know if if Tesla have implemented any SNNs as hydra "heads"?

 

[ZeApelido]

This quote caught my eye:

" The biggest, however, is that supervised machine learning doesn’t live up to the hype."

Ouch. This sounds like bad news for Elon's lofty FSD promises. Frankly, the fact that Elon has been promising L5 robotaxis since 2016 and we've yet to get "city NOA" should be a hint that perhaps the camera vision + supervised machine learning approach is not as easy as Elon thinks.

I am not saying that Tesla can't do good stuff with supervised machine learning. They have certainly made a lot of progress with AP since 2016. I am sure Tesla will do more good stuff. So supervised machine learning will do good stuff. I am not bashing AP and I am not saying that Tesla can't do anything. All I am saying is that it will take Tesla longer to achieve L5 autonomy than Elon thinks. Supervised machine learning is just one of many tools. It is not some magic bullet.

LOL shocked to hear a company that has scant access to real data say their supervised learning wasn't good enough.

Supervised learning will be good enough for Tesla I bet. They are probably the only company that can get enough relevant edge cases.

Tesla's problem will be if they can solve perception - can they harness 6 seconds of video and have enough training compute to get that much data to converge? Then if they can, what amount of processing power will be needed on each car to make inferences? Is it anywhere near the hardware they have?

 

[Daniel in SD]

LOL shocked to hear a company that has scant access to real data say their supervised learning wasn't good enough.

Supervised learning will be good enough for Tesla I bet. They are probably the only company that can get enough relevant edge cases.

Tesla's problem will be if they can solve perception - can they harness 6 seconds of video and have enough training compute to get that much data to converge? Then if they can, what amount of processing power will be needed on each car to make inferences? Is it anywhere near the hardware they have?

Why not attach a camera and microphone to 1 million people and use supervised learning to make an artificial human? That seems like it would be way more useful than a self driving car.

 

[diplomat33]

LOL shocked to hear a company that has scant access to real data say their supervised learning wasn't good enough.

Supervised learning will be good enough for Tesla I bet. They are probably the only company that can get enough relevant edge cases.

Tesla's problem will be if they can solve perception - can they harness 6 seconds of video and have enough training compute to get that much data to converge? Then if they can, what amount of processing power will be needed on each car to make inferences? Is it anywhere near the hardware they have?

Let's look at the full quote in context:

" The biggest, however, is that supervised machine learning doesn’t live up to the hype. It isn’t actual artificial intelligence akin to C-3PO, it’s a sophisticated pattern-matching tool."

I am not sure that the problem is the quantity of data. Rather, the quote is saying that people hype up supervised machine learning like it is AI and supervised machine learning can automatically teach a car how to drive with enough data. But they are saying "no", supervised machine learning is not AI, but rather it's just a sophisticated tool for matching patterns. So they are saying that even with enough data, supervised machine learning does not do as much as some people think it does.

 

[ZeApelido]

In no way did I imply supervised learning will get you AGI. That's laughable and as a machine learning scientist of course I believe supervised learning is really just pattern matching.

My opinion is that the act of driving will not require general intelligence, just like AlphaGo and all those reinforcement algorithms can sufficiently play many games without having AGI.

I don't believe learning how to drive given a known map of surroundings is going to be the limiting factor. Many more complex games have been beaten by computers.

Do you think a reinforcement learning algorithm can't be built to beat a driving simulation program? I think it can.

However getting the map of surroundings is a much harder problem. Is that a lead pipe on the road or a piece of paper. Is that person walking thinking about crossing it, or looking at their phone?

These pieces have to be really accurate. #1 you need a lot of edge cases to learn from. #2 you need a lot of compute power and right hierarchy to train the model (video is not easy).

 

[diplomat33]

Do you think a reinforcement learning algorithm can't be built to beat a driving simulation program? I think it can.

Yes, I think a reinforcement learning algorithm could be built to beat a driving simulation program.

In fact this video shows an AI use deep reinforcement learning to successfully auto park a car in a 3D simulation.

I am not saying it would be easy but I think it could be done.