What Tesla is doing, the data collection and algorithm optimization, is basically what nVidia and Google describe as deep learning.

Again, totally wrong. But you barely understand their business model so w/e.

ETA: I'd love to hear an explanation of what you think "deep learning" is. Should be a laugh.

Except they have by far the best production variant already deployed in 100k vehicles. For all the idiocy around the business of Tesla, they have some actual stuff that works and works great. Whether it can bail out a company spewing a billion a year is an entirely new question, but they do have some assets that could be valuable even in bankruptcy.

Deep learning is generally defined as something that has a neural network with multiple hidden layers.

We don't really know what Tesla are doing, but they are using Mobileye chips with custom software, chips designed for evaluating neural networks and they seem to rely on a camera based positioning, for cameras deep learning algorithms generally perform well. So likely they are using at least deep learning online. What they do offline and use for control we don't know, but I would guess that so far it's not that much deep learning, probably mostly classical algorithms such as particle- and kalman filters, graph based maps, sliding mode control/model predictive control, LQR etc rather than convolutional neural networks, reinforcement learning etc. But I would guess they are looking at the latter methods to potentially use them more in the future.

We really need to stop talking about Tesla and the autonomous driving end game as if they're anything other than a clown show. Tesla don't have the money to work on this, for a start. Tens of billions in research money will be flowing into this as it gets closer to reality. The major car companies and others who want a piece of this pie are capitalized in the trillions, collectively, with large positive cash flow.

Tesla are broke, and barely have enough money to survive to ramp up car production to produce 300K cars/year in 2020, let alone the billions to be spent on maxing out the "gigafactory" so their economics can work out. It's just absurd that people think Tesla's "research" in this area is anything but a stupid gimmick.

Anyone who thinks Tesla has a chance in this field is out of touch with reality.

I think you are making a generalization. Well you are. I'd say 99% of people who use deep learning mean "something something that I don't understand but want to sound smart about".

Regardless, parameterized edge cases create a virtuous cycle that further increases product lead (and deprives others of information). But I wouldn't really expect TS to know that.

Aside, last I checked ANN had massive issues with training data.

Again, you have no idea how this works. You deny the factual reality of them being the world leader in the field. (And weirdly neglect to mention Baidu which poached the premier Google talent, but I guess that doesn't fit your totally crazy narrative.) Beyond this, you lack the technical knowledge to understand how this could give an edge. You are a person who understands the carriage business and is seeing a car for the first time. Again, lets bet on how many mp are available on the first road approved SDC. O/U 150mp. I'll take as much as you have to offer.

Tesla is broke, but they have a legit shot at SDC. Remember your completely wrong prediction about their autopilot technology?

You don't need a hundred megapixel camera, nor a higher bandwidth cpu, nor deep learning, to recognize the objects on the road automatically because the tech to do that is already in use in autonomous car software:

9:40 in this vid for brief examples, 10:21 shows the captured bits and forms processed by machine learning (not deep learning), 11:11 for a funny example

Haha, true. But I still find that the industry generally is very sceptical to deep learning, very few seems to understand what the fact that RNNs are turing complete implies.

As for the training I would say that we have gotten pretty damn good at training these networks once there is good labeled data. Sure you need some GPUs and you need time, but buy that new card from nVidia for $129k, give it some data and wait some hours and you have gotten pretty far.

Disclaimer: I own NVDA stocks.

Because many of the more "advanced" methodologies are actually inferior?

I agree with what you're saying, but finding enough people in blue shirts who aren't police officers, people who are police officers and then people in gimmick police officers is a training set that is a massive PITA to gather. But when you have 100k+ vehicles gathering this data for you and being alerted to what problems they're having, you have access to data that is virtually impossible to acquire otherwise. (Obv that example sucks, but without actually deploying an AI, you can't really improve it since you have nfi how it recognizes things, esp in multi-layered NN).

The algorithms and hardware are basically in place.

I am actually inclined to agree with TS' assertion that Tesla's using outdated technology. They are using more or less off the shelf components, some of which were not originally meant for autopilot. The thing is I don't think it matters.

Data collection is the hard part. Nobody else has the kind of customer base eager and willing to pay six figures to become beta testers/lab rats. Like Mihkel05 and heltok pointed out, "self learning" algorithms are already pretty good if you have "good labelled data."

Tesla is going to have tens of millions of hours of "good labelled data" with how drivers react to various inputs. I wouldn't be surprised if they are already getting data on how people drive without auto-pilot.

Tesla are using pretty much the same as most of the industry right now. They do some custom software and have some extra communication, but the cameras and radar I assume are the same.

The big questions is how far you can get with current sensors and computing hardware and how far you can get with the next level. It seem that Elon thinks they can get pretty far.

My opinion on the topic is that in theory you should be able to get very far. We might not have the software and data for it now, that is a practical issue. But the cameras, radars and sonars imo see enough and the neural network should be able to be optimized to run on current hardware. The main issue imo is training the networks in the first place, not fitting it to the current hardware.

This is not something I really understand that well but when we're talking about exception handling that is critical to driving safely isn't processing power relative to needed speed of decision going to be a significant hurdle? All this is going to be have to done in car I would imagine so that seems like a significant road block that no one has addressed.

Otherwise you're trying to use a limited set of variables (object rapidly approaching) and creating a probability of a drastic reaction vs. the risk of a drastic reaction. So even if it does get to the point of being safer than a human driver it's going to be inordinately slow (say slowing down any time someone is semi-tail gating due to higher probability of drastic reactions), have to be turned off in rain, etc..

I think everyone here who should have an opinion on AI sees this issue the same way essentially.

To quibble a bit, computing power and algorithmic efficiency are just two sides of the same coin. There are many things no one really bothers to optimize anymore because hardware is super cheap. We don't need both.

Also, the cameras on the cars are extremely old. We have 8k video cameras. No one uses them. If this mattered, we'd see every robotic vision company paired with a camera company to be on the cutting edge of tech.

Why do you think a computer would be slower than a human with a limited set of variables? Computers are incredibly good at what you described.

I'm saying they will be forced to use a limited set of variables due to the limitations of onboard processing power and speculating that limited variables will never correctly account for the edge case scenarios that are so important in free range driving.

Again, talking out my ass but I think it's much more likely to have something like beacons in a defined range (city) that allow for automated driving (where the processing power can be offloaded).

The whole point of what I just wrote is that more data allows better algorithms to make decisions in more computationally efficient ways that handle edge cases better.

Also, the idea that you can offload processing like that is totally absurd. Like just a complete lack of understanding of how computers and networking is handled/works. You've asked me to provide an explanation, but what you wrote is like suggesting if we all flap our arms fast enough we can fly.

I'll admit I'm not sure what you're talking about but you can offload processing by sending data to a separate unit that has more processing power than can be contained in a car. For example, in my city scenario you could simply have a beacon in the self driving car that notifies the locally based camera/processing setup that it's there and then all the data and processing is handled outside of the car.

You can see simple examples of things like this with Google personalization of search results, data is captured by your browser but processed off your device.

Very much agree that better data drives efficiency but data isn't the only bottleneck in this scenario. I don't really know but it seems like given the amount of variables and the speed necessary to handle better than humans and the comparable problems that have been solved and how much processing power they took, processing is very a real bottleneck that doesn't have a clear solution besides offloading. Even offloading seems pretty speculative at this point given the speed necessary to make decisions relative to comparable problems that have been solved and the processing power they used.

The lag of online processing is too large.

There is "deep learning" to do with online processing but the decision making will almost certainly be local.

To put in chess/go AI terms, the cloud could pool together every game played globally and come up with the "best" parameters that is most likely to make best decision but the car has to make the decision locally once it is updated with the parameters.

Sure it will be "limited" to like 10M inputs at 100Hz. Paring this to a 13layer neural network(like the one that google used in AlphaGo to beat one of best humans at the game) would likely not be a challenge after optimizations.



It is running Alexnet(the one that was able to perform image classification on human level) at 2800pictures/second. And this is before optimization.

The first two bolded items are accurate. The third is a steelhouse style conjecture. I find it really ****ing funny you PMed me with that you knew about this. Since you don't know really anything about the subject. I'll bold the most relevant parts of the replies your received to help enlighten.

Let me know once you figure out what the processing bottlenecks are in terms of computational time with modern technology and then the transfer/receipt time utilizing modern network technology. I assume anything over 200ms is entirely useless, and that the development of on the fly processing is damn near impossible.

But hey "You have some expertise". Best of luck learning in the future.

heltok,

This is entirely off topic, but do you find the sort of video you linked really dumb/patronizing? When I watched the first 90 seconds, I listened to a guy recite slides over and over, then tell me why they were doing stuff which was completely inane.

One of the worst 90s of presentation I've ever seen. I assume it is gewgle technology that pwns, but if he is just gonna read it... slideshare plz?

I have seen worse. But sure, he is no Elon or Steve...

Also lolwtf:

Clearly they are moving into selling interface skins for $10 a pop.

Completely agree with this, Elon Musk actually said that he expected Tesla to fail at the start, which it should of. It's just one of his gimmicky side projects that for some reason people ploughed money into... because "omg, look this is revolutionary omfg!?{"

Now it's so big he doesn't have a clue what to do. He can't produce anywhere near enough units for the Model 3, having only 325k pre orders is completely irrelevant. Once it dawns on people that they have to wait over a year or two for a car, they'll soon be pulling out.

Yeah because a company that has a full order book will be on the verge of bankruptcy unlike a great company like GM that needs to be bailed out.

Your logic makes as much sense as "which it should of."