SASSquatch
Well-Known Member
- Joined
- Feb 3, 2022
- Threads
- 34
- Messages
- 1,842
- Reaction score
- 3,572
- Location
- Washington DC
- Vehicles
- BMW i3s Ford C-Max Hybrid
- Occupation
- Semi-Autonomous Yeti
You hit the nail on the head. The difficulty with Machine Learning (ML) is that you typically need exorbitant amounts of data. ML models need robust data points of every possible scenario and they are very susceptible to noise from degraded data signals.Great post!
Unlike @thrill I do not have formal experience with the technical details behind machine learning, but I have some related experience and classes. I know to never underestimate technology, but I think safe and reliable self-driving cars are still a ways off. I hope it happens, it could just be a while.
My armchair thinking is that a human being driving a car applies not just "car-driving" knowledge while at the wheel, but also all the other life-experience of living in this world. Maybe that knowledge isn't tapped when the driving situation is "standard": well defined lanes, well-behaved traffic, no random non-vehicle objects or dynamics in the environment. But "standard" is in quotes because how much of the time is real driving in that kind of environment? Maybe mostly for some, very little for others.
Training a car to drive by collecting data about a road and other cars is necessary, but not sufficient. What about how kids chase after balls or ride bikes? What about how tree branches appear in different light and different seasons? Is that light through the mist and rain at dusk a porchlight or another car with a headlight out? It's going to take a long time for self-driving algorithms that just use data from driving to back fill all those other life experiences that we apply when driving. Again, my speculation.
Plus, the computational capacity of the human brain is . . . impressive. Getting a silicon equivalent in a car will take a while.
In any case, I am thankful for this thread, and related threads, as a heads up. My car is old enough that even the most basic driver assistance features didn't exist when it was made. I didn't plan on using anything from the Rivian in this regard, and now for sure I won't.
When you are selecting "I'm not a robot" when you do a CAPTCHA test for some sites you will be asked to pick out a "fill in the bank" object from a series of pictures. But the pictures typically are blurry or grainy images.
The human brain is exceptionally good at object recognition (even severely degraded images) and generalizing to other exemplars with very few data points. You don't need to see every possible example of an apple to distinguish it from an orange, but a ML model might.
All the things human vision is very good at - seeing and recognizing a shadow or a tree branch vs something that would cause serious damage if you struck it, ML and AI models struggle with. There is still a very wide gap here.
A great example comes from how are visual system works. The fovea, the center of the retina which has the highest visual acuity, is packed with photoreceptors called cones that provide color vision in light conditions. Rods line the outer or peripheral area of the retina and are used for low light/non color vision.
When you are walking around at night in low light conditions, the fovea is one giant scotoma or blindspot but you don't perceive a giant hole in your vision. Why? Because your brain is literally filling in everything in that blindspot with what it is seeing in the periphery and what it is expecting to see and it's damn good at it because if you hadn't just read what I wrote (or are really up to speed on the human vision literature) you would have never known your visual system was doing that.
That is next level amazing that AI and ML is nowhere close to replicating.
My personal bias is that we can augment the human with driving aids but that the human should always be in control of the system - until at least we remove humans from the equation all together. But that is for another thread...
Sponsored
Last edited: