“Brittle AI” surfaces in vision-based training systems

[rr6013]

USAF target-recognition test reveals AI self-reported 90% accuracy verification only valid 25% in a cross-check of sensor-based AI.

Driverless car settings are referenced. It fortells that unless the training data includes enough angles, perspectives and instances for the explicit target(label in automotive nomenclature) the AI risks becoming over-trained on too high data confidence.

For one car example, LH Turn on two lane roadway at an intersection. Check oncoming traffic, check traffic behind is stopped, check clear intersection then proceed to LH Turn. What is wrong with this scenario? If data training that DOJO feeds FSD repeats instances of LH Turns of this type then it will complete 99% safely.

But FSD will miss the impatient driver who sees a line of cars slowed to a stop and chooses to enter the oncoming traffic lane to pass those stopped cars. FSD would execute LH Turn only to be T-boned by a car traveling in the wrong direction from behind while it tried to pass. This actually happened which my bad, I didn’t check left over my shoulder to make sure there was no oncoming traffic traveling the wrong direction. I only checked the mirror that traffic behind had slowed. That makes me “brittle”.

SO sure hope FSD is simulating traffic accident reports rather than the millions of accident-free instances of edge and corner cases that exist from Teslas. I’m not certain if cars behind a Tesla obscure rearward vision. THUS exclude PUREvision reading underneath a car for traffic from rearward direction. Brittle is a thing. It can happen to data…even you! LOL

Sponsored

 

[Buster Highman]

Pretty sure they are referring to CV that processes two-dimensional data. This is why Tesla Vision extracts 3D point cloud data from the cameras, allowing it to used 3D labeled classes such that it can identify the object from any angle.

 

[Crissa]

And a driver that darts across an intersection is still liable. The AI has no reason to expect or be liable to react to things it can't see. We don't expect human drivers to, why would we expect AI drivers to be able to guess that someone is about to violate traffic laws?

-Crissa

 

[ricinro]

And a driver that darts across an intersection is still liable. The AI has no reason to expect or be liable to react to things it can't see. We don't expect human drivers to, why would we expect AI drivers to be able to guess that someone is about to violate traffic laws?

-Crissa

As a human there are many times I anticipate human behavior including humans doing stupid things including speeding up and running red lights at nearly every intersection every day, not using turn signals, cutting corners, road rage, and distracted drivers.
Reading the road is psychology as well as recognizing the environment.
I listen closely when people describe FSD as akin to a student driver.

 

[PilotPete]

I listen closely when people describe FSD as akin to a student driver.

What if the student driver could learn from 5,000,000 other drivers submitting up to 19TB of video EVERY day? How long before the student becomes the master? (play Karate Kid music here)

Human action isn't random. People do what is in their best interest and generally, usually, most often don't fully evaluate the entire situation before acting. Example, on a bike and on a motorcycle you learn to watch parked cars on the side of the road. You look at the driver's seat for occupation, look at the mirrors for someone pretending to glance at the mirror or over their shoulder. And you EXPECT them to pull out in front of you. When you see someone that appears to intentionally drift to your side of their lane, you expect then to change lanes (signal optional). Since AI is watching the videos without preconceived ideas and just behavior, it will do a FAR better job of picking up these subtle hints and figuring them out. It will do better than any of us at anticipating behavior in the moments before they occur. And this moments WILL prevent accidents.

In the not too distant future, expect the deductible on your insurance to be one rate with FSD, and a higher rate when you are driving. Not 2024, but by 2034.

 

[SlegMD]

FSD will be a thing when all vehicles have FSD and communicate with one another seamlessly.

 

[ricinro]

FSD will be a thing when all vehicles have FSD and communicate with one another seamlessly.

I think this is where FSD will work; when humans can't interfere.

 

[PilotPete]

FSD will be a thing when all vehicles have FSD and communicate with one another seamlessly.

I think this is where FSD will work; when humans can't interfere.

It will be far better without human interference. However, you appear to think driving is a lot harder than it really is. FSD 11.X with 300K lines of code is a small app, not an enterprise level program. Now with AI, they're down to 3k with the AI part separate. Look, if even the below average people around us can drive (using only vision), so can the car. It ain't rocket surgery!

Now, EVERY time an FSD car gets into an accident, the humans will blame FSD. And the evening news will make a big stink about the FSD accidents. But remember, when something becomes common, the news quits reporting on it. So as long as it is so rare that you read about it online and the local news covers it, it's pretty rare. There are thousands of fender benders in southern California every morning and evening. And unless they involve a dozen or more cars and it closes a freeway, the news couldn't care less.

Sponsored