Wavye Teaching a Car to Drive

FrgMstr

Just Plain Mean
Staff member
2FA
Joined
May 18, 1997
Messages
58,702
This is an interesting and novel approach to self driving cars that is very different from anything else that is being used. Wayve is "simply" allowing onboard safety drivers to teach the cars how to driving using onboard AI, instead of using a series of maps, exotic sensors, and 10s of millions of parameters coded. Thanks cageymaru.

Check out the video.

Our network architecture was a deep network with 4 convolutional layers and 3 fully connected layers with a total of just under 10k parameters. For comparison, state of the art image classification architectures have 10s of millions of parameters.

Wayve has a philosophy that to build robotic intelligence we do not need massive models, fancy sensors and endless data. What we need is a clever training process that learns rapidly and efficiently, like in our video above. Hand-engineered approaches to the self-driving problem have reached an unsatisfactory glass ceiling in performance. Wayve is attempting to unlock autonomous driving capabilities with smarter machine learning.

Imagine deploying a fleet of autonomous cars, with a driving algorithm which initially is 95% the quality of a human driver. Such a system would not be wobbly like the randomly initialised model in our demonstration video, but rather would be almost capable of dealing with traffic lights, roundabouts, intersections, etc. After a full day of driving and on-line improvement from human-safety driver take over, perhaps the system would improve to 96%. After a week, 98%. After a month, 99%. After a few months, the system may be super-human, having benefited from the feedback of many different safety drivers.
 
On the other hand, i bet this system will crumble or be extremely prone to introducing (unknown) side effects when you wish to include some hard coded rules instead of using the organic AI driven learning.

Like a hard coded rule, if you see ball go across street, you slam on brakes, as there is high chance of child following it...
 
Now that's machine learning. Pretty cool - has the potential to greatly simplify things - it'll be neat to see what happens as it continues to improve.... especially if the cars themselves can communicate with a.. greater AI to "teach" other cars what they've learned? lol..
 
"Our network architecture was a deep network with 4 convolutional layers and 3 fully connected layers with a total of just under 10k parameters."

Whatever the F* that means, it still won't work.

Getting real tired of all the "neural networks" and "AI" that are just some algorithms that do nothing useful.
 
"Our network architecture was a deep network with 4 convolutional layers and 3 fully connected layers with a total of just under 10k parameters."

Whatever the F* that means, it still won't work.

Getting real tired of all the "neural networks" and "AI" that are just some algorithms that do nothing useful.

Ai (as in Aye!); whenever they say things like "Imagine deplo..." uh huh, I pretend lots of things all the time, especially in bed. "which is initially 95% the qual"...NO. I mean, there are a TON of humans that should not be driving and for sure I might feel safer - almost - with AI behind it. But, look at Facefail's algorithms for censoring posts, or Twitters, or Tesla's self driving modes, and those are really good minus the morons sleeping at the wheel and overriding the system. Still not good enough.

Really cool though, keep innovating peeps!
 
This is just the most basic supervised neural network training with backpropagation and some motorized interfacing. But credit where credit is due, they probably raised a small fortune from idiots. ;)

Consider what "the other guys" are doing: pitting TWO reasoning networks (CNN + RN) against eachother to check their work. Input is not only from the road reading sensors, but also from estimated futures based on incoming sensor data. So the AI effectively has realtime perception and realistic prediction to imagine possible outcomes and make decisions toward matching future incoming sensor data to the most ideal imagined outcome -- all based on strict boundaries. The more robust the boundaries, the less training is required, and that's the whole point of neural networks -- it helps to "discover" all the things you took for granted with experience and intuition and apply it algorithmically.
 
This is a completely worthless exercise in futility. This is not self driving, this is not intelligence, this is trial and error. But in traffic you're not allowed to make errors because any error could be your last.
this method of teaching an ai is perfect for simple jobs like the job of a robot on an assembly line, where there aren't many variables, and if it can do the task once it will continue to do it forever.

But in traffic there are millions of unforseen variables. If each time when a new variable is presented the system makes a mistake then it is completely worthless for driving.

Another snakeoil selling company trying to lure in ignorant investors. Do they have a kickstarter already? Or are they after bigger fish?
 
  • Like
Reactions: PaulP
like this
Meh. Sounds like they are in need of another round of funding and are wetting the path with a lot of marketing.
 
In my first CS masters class on video game programming (a good long time ago) we did an individual or group, semesters long project on pretty much anything we wanted to. I did mine on real time rendering of interactive particle physics (that was fun - and recreating physics isn't lol. Never got it working right) and there was a group that did it on machine learning with a simple AI playing an Asteroids type game (though with another ship controlled by the AI along with the Asteroids that you had to dodge). It was cool to watch. I can't remember how many rounds it took for the AI to learn how to dodge the Asteroids AND successfully attack the player, but I think it was about 10 rounds. So you'd try something else, the AI would learn that and after a few more rounds the game became impossible because it didn't take long for it to figure out how to out think the player (while dodging Asteroids). Add a second player, and it would learn that person even faster (maybe half as many rounds as the first player). The more players the more it learned. Eventually it didn't matter who played, the AI would win the game practically immediately.

Now, this was a 'simple' (didn't think it was simple at all at the time) machine learning AI running on some seriously old computers (based on todays standards at least) playing a game and it was impressive and repeatable.

It'll take a good while for it to "learn" to drive safely (and match current tech), but what I find especially amazing, because there WILL be accidents... is that the AI (if it's connected to some larger network) itself can actually learn from it. I wouldn't downplay this using modern tech specifically designed for machine learning and... sounds more promising than the current self-driving stuff. :)
 
In my first CS masters class on video game programming (a good long time ago) we did an individual or group, semesters long project on pretty much anything we wanted to. I did mine on real time rendering of interactive particle physics (that was fun - and recreating physics isn't lol. Never got it working right) and there was a group that did it on machine learning with a simple AI playing an Asteroids type game (though with another ship controlled by the AI along with the Asteroids that you had to dodge). It was cool to watch. I can't remember how many rounds it took for the AI to learn how to dodge the Asteroids AND successfully attack the player, but I think it was about 10 rounds. So you'd try something else, the AI would learn that and after a few more rounds the game became impossible because it didn't take long for it to figure out how to out think the player (while dodging Asteroids). Add a second player, and it would learn that person even faster (maybe half as many rounds as the first player). The more players the more it learned. Eventually it didn't matter who played, the AI would win the game practically immediately.

Now, this was a 'simple' (didn't think it was simple at all at the time) machine learning AI running on some seriously old computers (based on todays standards at least) playing a game and it was impressive and repeatable.

It'll take a good while for it to "learn" to drive safely (and match current tech), but what I find especially amazing, because there WILL be accidents... is that the AI (if it's connected to some larger network) itself can actually learn from it. I wouldn't downplay this using modern tech specifically designed for machine learning and... sounds more promising than the current self-driving stuff. :)
I think the lack of confidence comes from the fact that this demonstration is no different from your classmates project. We should be well beyond being impressed by neural network demonstrations as unsupervised control theory isn't new, but lots of young engineers are treating it like it is and using impressionable ignorance to garner investment attention. Control theory goes much deeper than throwing a NN at a problem and watching it learn. "Wee it spins!" is not the point.

For decades realworld, mission critical applications of neural networks required a service level proof before implementation could be considered. That's too difficult to do with meager implementations of single neural networks, like a BPNN. The system needs to not only replace portions of itself with the best iteration of an unproven algorithm (standard NN stuff), but also generate provable, exportable algorithms for review -- these days, using additional neural nets. e.g. Current "self-driving stuff" is actually well beyond this stage and uses many concurrent, realtime NNs of various types to self analyze and discover new rules that must be applied to the net which we hadn't thought of before. That's the process of automated proof, which is something more deserving of investment attention.
 
*shrugs* - The semester project (with a couple (or few?) students working on it with no previous experience) were able to do something cool well over a decade ago. Granted, my more personal experience on this stuff ended not long after that, but I'm thinking that a group of professionals working on this (?) using current tech might have better results.. I mean, I'm sure that the research knowledge and tech involved has improved at least a little in the past ten years. right? I'm all for trying out something new and see if it succeeds or, well, crashes. :)
 
The biggest problem is that even the most complex "deep learning" system has no knowledge of the world, and no "common sense". So if it comes across something new it will not know what to do, even if it is obvious to a human, like you don't have to brake for a tumble-weed blowing across the road. And there are almost an infinite number of new situations that could confront it, way too many for the "expert" drivers to teach it. But most of these can be handled by humans because we KNOW a lot about the world around us and can extrapolate from similar experiences. These neural-net expert systems can't do that, so they will always be doing stupid things and you will never be able to predict when this stupidity is going to manifest since the network structure is a complete mystery.
 
The biggest problem is that even the most complex "deep learning" system has no knowledge of the world, and no "common sense". So if it comes across something new it will not know what to do, even if it is obvious to a human, like you don't have to brake for a tumble-weed blowing across the road. And there are almost an infinite number of new situations that could confront it, way too many for the "expert" drivers to teach it. But most of these can be handled by humans because we KNOW a lot about the world around us and can extrapolate from similar experiences. These neural-net expert systems can't do that, so they will always be doing stupid things and you will never be able to predict when this stupidity is going to manifest since the network structure is a complete mystery.
Good decisions can be based even on imagined experiences. Reasoning algorithms can operate on imagined future states and learn from them in realtime all the same. Common sense is definitely not a problem. The real problem is jumping the gun and running cheesy experiments using elementary neural nets, and failing to even try implementing a proper system which can reason about good decisions and proove that they are good decisions.
 
Good decisions can be based even on imagined experiences. Reasoning algorithms can operate on imagined future states and learn from them in realtime all the same. Common sense is definitely not a problem. The real problem is jumping the gun and running cheesy experiments using elementary neural nets, and failing to even try implementing a proper system which can reason about good decisions and proove that they are good decisions.
I disagree; common sense is the big problem because it depends on having a massive cross-linked database of facts about the world and how it works. Current "AI" systems are a long way from having that.
 
I disagree; common sense is the big problem because it depends on having a massive cross-linked database of facts about the world and how it works. Current "AI" systems are a long way from having that.
Right, but that's the point of using neural nets in the first place. It's so that you don't have to explicitly detail every logical interaction of every nuanced attribute of the world. Instead, reasoning networks can derive and prove ideal solutions and the complexity of said solutions for each derived attribute. A similar process is solving logic problems with Karnaugh maps, or using automated proofing languages to write declarative code which can then define the imperative code without error.
 
Right, but that's the point of using neural nets in the first place. It's so that you don't have to explicitly detail every logical interaction of every nuanced attribute of the world. Instead, reasoning networks can derive and prove ideal solutions and the complexity of said solutions for each derived attribute. A similar process is solving logic problems with Karnaugh maps, or using automated proofing languages to write declarative code which can then define the imperative code without error.
And a neural net complex enough to be able to handle the unconstrained world outside the lab will have to be as complex as the human brain and will take years to train. We already have that; they are called children.
 
Back
Top