From what I understand, the output from programs using a machine learning algorithm continuously change as more inputs change the variable weightings of whatever model they have.
But the lane and road edge detection thresholds for my car’s sensors were likely determined (i.e, they picked and set fixed averages) from lots of data and set them as static values in the program to compare against. That’s just using data and statistics as usual.
There are toy cars for students of programming to write simple self-driving algorithms from sensor data to drive without colliding into walls and black or colored tape. Same idea. But I wouldn’t call it “AI,” which is a marketing term to begin with, not a scientific one.
Anyway, certainly no ML algorithm was running on the car’s system. The driving assist was very simple and predictable, and it was manufactured in 2018.
Generally ML models are trained once and then used in their trained state unchanging. They don’t continue to learn.
I went down a slight rabbit hole on this. As far as i can tell (looking at manufacturer claims on their website for devices of the era), 2018 is right on the edge of when a lane finding algorithm would have started to incorporate ML techniques. That’s when mobileye released the eyeq4, which claimed to run ML models, and Wikipedia says was used by Ford, BMW, Nissan, VW, and Honda, albeit on only some of their models. The previous Eyeq3 from 2014 seems to be more classical models, but still used some ML techniques, seemingly more for things like speed limit reading.
ML algorithms refer to quite a wide variety of approaches. Generally what separates them is when the parameters of a generic model/algorithm are optimized over some specific data, instead of using an algorithm handcrafted to the particular problem.
From what I understand, the output from programs using a machine learning algorithm continuously change as more inputs change the variable weightings of whatever model they have.
But the lane and road edge detection thresholds for my car’s sensors were likely determined (i.e, they picked and set fixed averages) from lots of data and set them as static values in the program to compare against. That’s just using data and statistics as usual.
There are toy cars for students of programming to write simple self-driving algorithms from sensor data to drive without colliding into walls and black or colored tape. Same idea. But I wouldn’t call it “AI,” which is a marketing term to begin with, not a scientific one.
Anyway, certainly no ML algorithm was running on the car’s system. The driving assist was very simple and predictable, and it was manufactured in 2018.
Generally ML models are trained once and then used in their trained state unchanging. They don’t continue to learn.
I went down a slight rabbit hole on this. As far as i can tell (looking at manufacturer claims on their website for devices of the era), 2018 is right on the edge of when a lane finding algorithm would have started to incorporate ML techniques. That’s when mobileye released the eyeq4, which claimed to run ML models, and Wikipedia says was used by Ford, BMW, Nissan, VW, and Honda, albeit on only some of their models. The previous Eyeq3 from 2014 seems to be more classical models, but still used some ML techniques, seemingly more for things like speed limit reading.
ML algorithms refer to quite a wide variety of approaches. Generally what separates them is when the parameters of a generic model/algorithm are optimized over some specific data, instead of using an algorithm handcrafted to the particular problem.