181–190 of 221 posts
Earlier quoted context omitted.
This is the golden rule of software dev too - promise features or promise a release date, but never both.
Be conservative in what will be delivered, and liberal with when it'll be delivered.
Earlier quoted context omitted.
every now and then I do something while driving that, IMO, requires "human cognition". Not just image recognition, obstacle avoidance, rule following, etc. Like, I'm cruising along at 45mph and a plastic bag is blown in front of my car. Yes, I see an obstacle and need to make a decision about what to do - but I do this as a human who knows what a plastic bag is, and how it is far safer to drive right over it than sla…
While that's true, the autonomous car can also react so much faster than you can. It may be able to largely trade-off your level of contextual awareness for pure reaction speed in most cases.
You need a certain hysteresis/threshold to filter out false positives.
Earlier quoted context omitted.
> Slamming on the brakes for a plastic bag is not the universal safer choice. ...I didn't say it was. And you call me the one making assumptions. > reaction/breaking time is meaningless without timely recognition. If it can't quickly recognize large moving objects then it's not qualified to drive at all . I'm assuming it can do that because self-driving at all is part of the premise. (Well, specifically the recogniti…
1) I'll rephrase: occasionally (and needlessly) stopping for a plastic bag (or any other false positive) should not be considered a massive improvement over current human abilities 2) >>>you might preemptively slow down, but a properly designed system can brake a whole lot faster. >If it can't quickly recognize large moving objects I'm not questioning whether it can recognize quickly. Just whether it can ever be bett…
Okay, but I didn't say that either. I said it would be an improvement to the experience I have. I will trade an abrupt stop in front of a plastic bag for several hours of being able to do what I want while the car does the driving.
> If AI can react and brake 2 seconds faster than a human, but doesn't recognize the threat until 4 seconds after humans can, then AI drivers will kill more people than human drivers. Which won't be fine.
The AI processing time to recognize things in a frame needs to match the framerate, or the system is just broken. Maybe it waits for 3 frames to show the same thing before acting, that's fine. That's still faster than a human. Especially if you use a 240Hz camera, which is not very expensive. And you can and should put in enough computing power to run the 'reacting' system at the same frequency.
Then once it decides to brake, instead of waiting for a foot to move it can put the virtual brake pedal at any position in under a millisecond.
I can imagine a car so poorly designed that its reactions are slower than a human. But it's hard to imagine a car that's simultaneously certified for letting me take my eyes off the road, and is that broken.
So I still say that a self-driving car would be able to brake quicker and harder, and it should be fine. Even without the ability to predict.
Earlier quoted context omitted.
Not sure there is enough evidence in support of those claims about Tesla. Tesla doesn't collect anywhere near enough data from its customer fleet to support the kind of massive training being claimed. In contrast Waymo and Cruise vehicles are regularly in the depot where each car offloads terabytes of data. Who has the bigger training set?
Tesla routinely gets data from the fleet and analyzes it. Here is Andre Karpathy speaking on this subject: https://youtu.be/Ucp0TTmvqOE 1:49:30. I’m on mobile otherwise I would link to the exact spot. Tesla has tens (hundreds?) of thousands of vehicles on the road right now that are continuously feeding training data back to Tesla. What evidence is there that this is insufficient, or even deficient? From what I can g…
I have very mixed feelings about autonomous cars. One side of me is so excited about the prospect and finds it all amazing and futuristic. I love reading about how Tesla and Waymo approach it differently. The other side of me is absolutely not ready to trust my life to a machine - even though I know I do that all the time in other ways in modern life. I think it will be people's emotional evaluation that will matter…
every now and then I do something while driving that, IMO, requires "human cognition". Not just image recognition, obstacle avoidance, rule following, etc. Like, I'm cruising along at 45mph and a plastic bag is blown in front of my car. Yes, I see an obstacle and need to make a decision about what to do - but I do this as a human who knows what a plastic bag is, and how it is far safer to drive right over it than sla…
Earlier quoted context omitted.
I think it's the other way around: Direct mapping of inputs to outputs will work great for a subset of conditions, but the last N% is infeasible without explicit logical programming.
You'll never run out of edge cases if you have to explicitly program the logic. At some point you have to let the machine learning decide it. Maybe there is room for a hybrid approach, but I think anything relying on programmers having thought of every situation ahead of time is doomed to fail.
I honestly don't expect go-anywhere L5 in the next decade at least, and I definitely expect an engineered approach to hit L4 first.
Earlier quoted context omitted.
Maybe semantics, but the biggest problem isn't necessarily trusting the machine to do what's expected. The problem is the machine being able to predict what pedestrians and other drivers are going to do. This isn't just a problem for driverless cars, it's become a real problem for people, too. The rules of the road are there for a reason, people! When you wave someone into traffic, or have your emergency flashers on…
> The problem is the machine being able to predict what pedestrians and other drivers are going to do. That's something machine learning can do pretty well, actually. There has been a lot of research on this topic, and not applied just to self driving - natural language generation, stock prediction, playing games with multiple agents, even content recommendation do future prediction based on past information.
I'm not familiar with the games, but from what I've seen with all the other things listed, that's not a technology that I'm anywhere near ready to trust my life with.
Earlier quoted context omitted.
I think that you overlook the fact that a tremendous amount of evolution went on before humans came onto the scene ~200k years ago.
I was waiting for that :). Let's put it out at billions of generations for pre-human evolution. Point stands. I computer will do millions of simulations in a day. And be done across a population far greater than historic evolution. Doesn't take long to get billions of learning events. On top of which evolution pre-intelligence (as a very loose term) was somewhat random, where our systems are starting from a much more…
Earlier quoted context omitted.
1) I'll rephrase: occasionally (and needlessly) stopping for a plastic bag (or any other false positive) should not be considered a massive improvement over current human abilities 2) >>>you might preemptively slow down, but a properly designed system can brake a whole lot faster. >If it can't quickly recognize large moving objects I'm not questioning whether it can recognize quickly. Just whether it can ever be bett…
> 1) I'll rephrase: occasionally (and needlessly) stopping for a plastic bag (or any other false positive) should not be considered a massive improvement over current human abilities Okay, but I didn't say that either. I said it would be an improvement to the experience I have . I will trade an abrupt stop in front of a plastic bag for several hours of being able to do what I want while the car does the driving. > If…
One word: Chauffeur
>The AI processing time to recognize things in a frame needs to match the framerate
If you're talking about framerates at all, you're still missing the point. I'm talking about human ability to predict and anticipate. The fact that a human can brake before there's anything in the frame to even detect. We've (humans) got a head start, so the speed of the AI isn't all that important.