by BeauV » Tue Oct 02, 2018 12:20 pm
OM,
I completely agree, with some of what you've said. Way back when Moses was in diapers I was on the board of a company (Quintus PROLOG) which attempted AI based on declarative statements of formal logic. For reasons that appear obvious now, those declarations were entirely insufficient for this sort of problem. Fortunately, no one believed that this sort of problem was even remotely possible (not really, we all talked about it a lot) so the only thing Quintus software got applied to was getting the green door on the green Buick as it went down the factory line. We were quite good at that.
The breakthrough came when we started training AI software with massive numbers of real-world experiences. I've fooled around a little with trying to predict hurricane tracks by simply training the software with every piece of data I can find for every hurricane we have seen. The darned thing got pretty damn good, but only about 80% of what a human aided by computers can do. I think my hurricane test is a pretty accurate assessment of where we are: Humans augmented by AI software are the best solution we've found. Stan has trained a system to find and predict weather patterns which will do things like sweep from NY to Ireland with one front. Thus, the repeated Transatlantic records.
Returning to your comment about resolving time/distance/intercept etc... in real time, there is obviously a terrific Darwinian force that would select for that. Hunting. Imagine what a peregrine falcon does in its brain as it dives towards a bird at over 200 MPH! I've watched the falcons near our old office nail a fast moving pigeon and have read that the visual system within a falcon can spot a small bird (dinner) from over 3/4 of a mile away, once it moves. That's a tip, to my mind, it moved. Note also that the peregrine has over 80% of its brain dedicated to its visual system, which makes sense when you're diving at a pigeon that is flying 20' above the ground and your'e going 200+ MPH straight down! Make a mistake and you hit the dirt, mistake the other way and you miss dinner. There is some evidence that the falcon estimates the weight of the bird it is about to strike and adjust speed accordingly, counting on the mass of the bird it will impact to slow its decent. Falcons can withstand over 30G impacts, which is how they actually kill their prey, they ram it. The talons are only there to keep the prey from falling to Earth after the strike. There are witnesses who have seen peregrine miss a strike and the falcon is often killed by the impact with the ground below. Darwin at work.
Back to humans v cars. I strongly believe that our visual systems are more highly tuned to hunting activities than we realize. This is why we can catch a baseball on the run, return a 100 MPH serve in tennis, etc... We are not only calculating the intercept, we're controlling our legs, hands, and everything else to deliver the death blow to that pesky tennis ball. I strongly believe that this is why we can drive cars, especially at high speeds, and why we are so inept at it when we've put our visual systems to work looking at a phone screen. To drive well, one must notice the movement of everything in the visual field and track the changes. Looking at a phone stops all that and it takes our visual systems numerous seconds, if not a minute, to catch up and re-establish situational awareness.
For these reasons, I think that driving a car well (which customers REALLY want us to do) will be one of the toughest tests for AI. Unlike financial calculations about stock prices or predicting the path of hurricanes, driving matches the computer up against one of the things which humans (and lots of other predators) do REALLY well. That's what makes this effort so interesting - it's really hard!
____________________
Beau - can be found at Four One Five - Two Six Nine - Four Five Eight Nine
