“Brittle AI” surfaces in vision-based training systems

rr6013

Well-known member
First Name
Rex
Joined
Apr 22, 2020
Threads
48
Messages
1,343
Reaction score
1,281
Location
Coronado Bay Panama
Website
shorttakes.substack.com
Vehicles
1997 Tahoe 2 door 4x4
Occupation
Retired software developer and heavy commercial design builder
Country flag
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

 
First Name
Kevin
Joined
Jul 15, 2021
Threads
0
Messages
6
Reaction score
41
Location
Fairbanks
Vehicles
Cybertruck
Occupation
Dr
Country flag
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

Well-known member
First Name
Crissa
Joined
Jul 8, 2020
Threads
90
Messages
9,977
Reaction score
16,129
Location
Santa Cruz
Vehicles
2014 Zero S, 2013 Mazda 3
Country flag
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

 

 
CYBERBACKPACK
Top