Neural Networks

[DeFries]

Everyone here seems very impressed by the latest update. My S is on 2018.12 5eadc71 - is that the latest? Reason I'm asking is that today I was driving on a 2x2-lane highway here in Netherlands and I was actually a bit disappointed with AP2(.5) behaviour.

The previous version seemed more solid; it solved the problem with crests in the road and kept the car centered quite well. The current version seems to be too relaxed in curves, letting the car drift to the outside of the curve quite far before correcting; it also doesn't center the car as rigorously, seems to avoid the verges and drive closer to the middle of the road. That just doesn't feel comfortable when overtaking large trucks or in traffic jams with motorbikes trying to cut through traffic by driving in-between lanes. I wish there was some way of telling AP to stay closer to the verges when in the outer lanes and not in a curve.

 

[atokosch]

Personally I only study the neural network related features. The driving state machine is pretty opaque and I think it's detailed behavior depends on a lot of external configuration parameters, map data and so forth. There are several other people on TMC who know a lot more about it - maybe one of them will respond to your question.

But I've seen similar changes in behavior on the same version of firmware and other people have reported them as well. I think the general sense is that it's the result of map updates, which happen continuously in the cloud, and of non-firmware configuration changes which get pushed to the cars without notification.

Ahh okay, I see. I wasn't sure if the neural network was able to be updated without a physical firmware update, but what you said makes sense. I have only had my Model 3 for 5 weeks now so I didn't know that updated maps via the cloud could help in the driving of the car, and from what I previously understood was that Autopilot didn't talk to/use any map data based on what Tesla has said after the Model X crash a few weeks ago. But if it's using some sort of data then that is cool.

Make sense for the driving state machine to be opaque and be controlled by external parameters (assuming you mean conditions, what it sees, etc) and map data.

I find your studying of the neural network interesting, and how autopilot works in general very interesting. This summer a few classmates and I are converting a Forumla Electric (FSAE) car to have autonomous capabilities. We are currently building the car as a proof of concept for our college so it isn't going to competition, and they want us to make it autonomous ready for a future senior design project. Unfortunately I think we will only implement line following but it will be a start.

 

[atokosch]

The new behavior is tied to the ADAS tiles and TomTom speed limit database which uses GPS to update speed limit information in the IC. The ADAS tiles have curve information for AP to know how to slow down.

You can read about it in the Tesla Autopilot Maps thread by @verygreen in this particular sub-forum. We are now locked out from accessing the tiles but it was cool seeing the data populate thanks to the awesome work by @DamianXVI

Oh thats awesome. Thanks for the info. I will check it out.

 

[malcolm]

Well, both time and space complexities seem to be improving:

Andrej Karpathy on Twitter

 

[hill]

Everyone here seems very impressed by the latest update .....snip.....

not every one .... others just read on, & think to their self, "yea sure" - as the X continues to dive towards the freeway off-ramp ... & still can't Lane change on regular boulevards.
Just go w/ it & let people continue being happy as clams.

.

 

[jimmy_d]

not every one .... others just read on, & think to their self, "yea sure" - as the X continues to dive towards the freeway off-ramp ... & still can't Lane change on regular boulevards.
Just go w/ it & let people continue being happy as clams.

.

Lane changing is blocked on surface streets now. It used to work.

 

[S100D]

Lane changing is blocked on surface streets now. It used to work.

Jimmy, lane change on surface streets always worked on my AP1 S85D but it never worked with my AP2 S90D. Do you have an AP1 car?

 

[croman]

Lane changing is blocked on surface streets now. It used to work.

It used to work with AP1. It never worked with AP2, only geofenced divided limited access highways

 

[jimmy_d]

Jimmy, lane change on surface streets always worked on my AP1 S85D but it never worked with my AP2 S90D. Do you have an AP1 car?

It used to work with AP1. It never worked with AP2, only geofenced divided limited access highways

Oh you're totally right. I was remembering that it worked on my AP1 car (I upgraded last September). I guess I'm not sure if it ever worked on my AP2 car.

But I am confident that the 'not working' is intentional.

 

[croman]

But I am confident that the 'not working' is intentional.

It definitely is. I wonder if they will activate local alc when the side repeater cams are active. I find ALC to be very smooth and able to navigate tight traffic now. Its incredible how refined the ultrasonics are after 2018.10.4

 

[CK_Stuggi]

Still has problems allocating cars to their proper lanes when merging from the left to the center lane. 80% of the time the vehicle breaks if a truck or other slow moving vehicle is in the right lane and between 75 and 200m (200 to 600ft) away.

 

[croman]

Still has problems allocating cars to their proper lanes when merging from the left to the center lane. 80% of the time the vehicle breaks if a truck or other slow moving vehicle is in the right lane and between 75 and 200m (200 to 600ft) away.

I no longer get that. 2018.12. It has been tracking vehicles that are lane changing across multiple lanes for me as I ALC. I just wish it displayed the vehicles in all lanes like AP1, so I could verify and gain confidence in the system. That information is irresponsibly being withheld as I believe the vehicle's object recognition is now superior to AP1 and it could likely display vehicles in all lanes (and perhaps classify them).

2018.5 did that but it was limited access and no further builds do it but I don't see why it couldn't. I notice the objects are internally present but only ALC brings that information to the user. That's not cool but I am glad the system is at least more capable than it usually reveals.

There is some adjacent lane car detection issue in it places the vehicle in my lane but it actually is in another lane. I notice the issue more on curves, but mine has improved greatly since 2018.10.4. 2018.12 seems to have improved in the 5 weeks I've had it (though I understand that is hotly debated, I do see improvements somehow and adjacent lane car positioning is one of them (particularly trucks which in 2018.10.4 would constantly cause me to unnecessarily slow as I passed them)).

 

[chispas]

I no longer get that. 2018.12. It has been tracking vehicles that are lane changing across multiple lanes for me as I ALC. I just wish it displayed the vehicles in all lanes like AP1, so I could verify and gain confidence in the system. That information is irresponsibly being withheld as I believe the vehicle's object recognition is now superior to AP1 and it could likely display vehicles in all lanes (and perhaps classify them).

There is some adjacent lane car detection issue in it places the vehicle in my lane but it actually is in another lane. I notice the issue more on curves, but mine has improved greatly since 2018.10.4. 2018.12 seems to have improved in the 5 weeks I've had it (though I understand that is hotly debated, I do see improvements somehow and adjacent lane car positioning is one of them (particularly trucks which in 2018.10.4 would constantly cause me to unnecessarily slow as I passed them)).

I’m on 2018.12 and I experienced breaking for vehicles on adjacent lanes that weren’t changing lanes or doing anything weird.

My car also renders the car in front of me outside the lanes in the screen if the car is more than 40 or 50 feet away, even in completely straight roads.

 

[MarcusMaximus]

It definitely is. I wonder if they will activate local alc when the side repeater cams are active. I find ALC to be very smooth and able to navigate tight traffic now. Its incredible how refined the ultrasonics are after 2018.10.4

On this topic, I did a trip down to San Diego and back(~400 miles each way) over the weekend. Had exactly 2 times where I had to disengage autopilot, and both were during auto lane change. The first was because, halfway through the lane change, the car suddenly turned sharply, moving most of the way into the lane past the one I was changing into before I could disengage and the second because it suddenly bailed on the lane change when almost finished and swerved back towards the lane I was in before. Still problems there at times, and I wouldn't feel comfortable given that using it on surface streets.

 

[jimmy_d]

On the topic of NNs:

AI and Compute

So the upper bound of compute intensity for training notable NNs has been going up at 10x per year for the last 6 years. The authors of this document expect that pace to continue for a while yet.

Google's recently announced TPU3 pods have a raw throughput of 100 petaflops (100,000,000,000,000 operations per second). To put that in perspective: various estimates of the computational performance of the human brain tend to fall into the range of 10^16 to 10^19 ops/second, so we are now in the zone (plus or minus a year) where cutting edge NN training hardware runs at a capacity roughly comparable to the human brain. And at the current growth rate in just a few years the best machines will be 100x the speed of the human brain, which is sort of the expected threshold where development of algorithms of human brain complexity can be done.

The next five years are going to be AMAZING.

 

[mongo]

Google's recently announced TPU3 pods have a raw throughput of 100 petaflops (100,000,000,000,000 operations per second).

Don't short change Goggle, It should be 100,000,000,000,000,000. (100*10^15 or 10^17).
The 2.0 TPU cluster was good for 180 petaflops, and they are coming out with a 3.0 version soon that is 8x faster...

 

[hill]

On this topic, I did a trip down to San Diego and back(~400 miles each way) over the weekend. Had exactly 2 times where I had to disengage autopilot, and both were during auto lane change. The first was because, halfway through the lane change, the car suddenly turned sharply, moving most of the way into the lane past the one I was changing into before I could disengage and the second because it suddenly bailed on the lane change when almost finished and swerved back towards the lane I was in before. Still problems there at times, and I wouldn't feel comfortable given that using it on surface streets.

yea ... that scenerio is still ongoing - & similarly beware, someone in front of you (going slower than your AP is set for) changes out of your lane - your tesla now speeds up to the flow of traffic - meanwhile in an adjacent lane, about 1½ car lengths in front of you a car signals to merge in front of you - begins to come over, even as you're speeding right into him. Thank god for human alert reactions - quicker than any alleged neural net. Otherwise you might have been creamed, like the inattentive X driver in the Cali Bay area. Yay/go neural net ... & yay smooth as silk.

.

 

[Pale_Rider]

That will require the fisheye camera to be used for object detection instead of just rain sensing.

Until then the car is blind in those areas as the air pressure is to great for the ultrasonics to be effective.

 

[jimmy_d]

Don't short change Goggle, It should be 100,000,000,000,000,000. (100*10^15 or 10^17).
The 2.0 TPU cluster was good for 180 petaflops, and they are coming out with a 3.0 version soon that is 8x faster...

doh.

and here I thought I was being hyperbolic by actually typing out the number and then went and undercounted the zeroes by 3.

thanks for the correction

 

[chispas]

This is what Karpathy tweeted earlier today

“lots of exciting recent work in large-scale distributed training of neural nets: (very) large-batch SGD, KFAC, ES, population-based training / ENAS, (online) distillation, ... ”

Maybe @jimmy_d can explain that to the mere mortals like myself.