TMC is an independent, primarily volunteer organization that relies on ad revenue to cover its operating costs. Please consider whitelisting TMC on your ad blocker and becoming a Supporting Member. For more info: Support TMC

Neural Networks

Discussion in 'Autonomous Vehicles' started by lunitiks, Nov 5, 2017.

  1. gregor

    gregor New Member

    Joined:
    Nov 8, 2017
    Messages:
    4
    Location:
    Switzerland
    Great.
    The videos shown or referred to do not seem to indicate that traffic lights are captured or interpreted.
    That is surprising. I can understand that this is not yet used for autonomous Tesla driving, as even the slightest error can have lethal effects, but a neuronal network should at leat start learning to tead traffic lights.
     
  2. mongo

    mongo Well-Known Member

    Joined:
    May 3, 2017
    Messages:
    8,898
    Location:
    Michigan
    Other possibilities: Data collected by NN is not necessarily all exported. It may also be a feature that is disabled on the HW 2.x platform due to processing limits.
     
    • Like x 2
  3. Cosmacelf

    Cosmacelf Well-Known Member

    Joined:
    Mar 6, 2013
    Messages:
    5,871
    Location:
    San Diego
    Acucuracy to 1 foot isn’t good enough, you are using localization as an additional data source for lane keeping, and figuring out where the turn in for turns. You need accuracy 10x better than that. And anyways, as I’ve said before and we’ve all witnessed, GPS isn’t infallible. It makes mistakes and isn’t always available (big tree cover, or surrounded by tall buildings).

    Localization by side of road signs is pretty cheap once you have the data. It doesn’t cost much in terms of local computing power.
     
    • Like x 1
  4. malcolm

    malcolm Active Member

    Joined:
    Nov 12, 2006
    Messages:
    2,914
    If it's not going to be reliable in city areas the greater accuracy seems kinda pointless.

    Too many systems giving different location info; cameras lidars and now GPS. Fine when they all agree, but when they don't, which one is telling the truth?

    Our natural assumption is the cameras, but only because we assume computers"see" what we do.
     
  5. Cosmacelf

    Cosmacelf Well-Known Member

    Joined:
    Mar 6, 2013
    Messages:
    5,871
    Location:
    San Diego
    You use statistical algorithms along with temporal smoothing to integrate different inputs. All sensors are "noisy" and each sensor can flake out. That's the whole point of redundant systems. Combining their outputs results in an overall more accurate picture of the world, but also a more reliable one.

    I had an experience recently which was a great example of how the human brain does this sensor fusion. I was trying to fit a key into a keyhole in very dark surroundings. When looking at the keyhole area it was full of visual noise since it was so dark. I couldn't make out what I was seeing. When by random fumbling, I got the key into the keyhole, all of a sudden, in a split second, I could "see" the keyhole and the surrounding area.

    What happened? My vision didn't suddenly improve. My vision system, however, got some extra information from the position of my hand and memory of what the keyhole should more or less look like. So it made an educated guess that this very noisy visual signal should fit the picture of a keyhole right THERE according to my proprioceptor network (position information).

    Sensor fusion is the next step is advanced AI, in essence, combining the outputs of several specialized neural nets together.
     
    • Like x 4
    • Informative x 2
    • Helpful x 1
  6. Snuffysasa

    Snuffysasa Member

    Joined:
    Feb 24, 2017
    Messages:
    257
    Location:
    Minneapolis
    1) Unless you don't need your application to work in cities, than its not pointeless.

    2) The localization methods that @Cosmacelf has been describing ARE reliable even in cities.

    I would answer this @Cosmacelf but already has.


    No not overkill. You need to listen to what @Cosmacelf has been saying
     
    • Like x 1
  7. WillK

    WillK Member

    Joined:
    Dec 8, 2016
    Messages:
    118
    Location:
    Milwaukee, Wi
    Any idea what changes they have made in 2018.39.x to the NN? I believe a major change was activating all 8 cameras. Wondering if anything else has been discovered.
    Thanks!
     
  8. Reciprocity

    Reciprocity Active Member

    Joined:
    Feb 27, 2017
    Messages:
    4,161
    Location:
    Chicagoland
    Lvl5.ai is an example of high def mapping purely from video feeds. In this case, dash mounted cell phone cams. 2-3cm is ideal but probably closer to 5-10cm or about the size of a business card. Just good enough for localization and as emergency backup in the case other parts of the system fails, the car can come to safe stop until a human driver can take over. I see high def maps as a redundancy in system without duplicates of every camera, radar and sonar.
     
    • Like x 2
  9. jimmy_d

    jimmy_d Deep Learning Dork

    Joined:
    Jan 26, 2016
    Messages:
    412
    Location:
    San Francisco, CA
    NN Changes in V9 (2018.39.7)

    Have not had much time to look at V9 yet, but I though I’d share some interesting preliminary analysis. Please note that network size estimates here are spreadsheet calculations derived from a large number of raw kernel specifications. I think they’re about right and I’ve checked them over quite carefully but it’s a lot of math and there might be some errors.

    First, some observations:

    Like V8 the V9 NN (neural net) system seems to consist of a set of what I call ‘camera networks’ which process camera output directly and a separate set of what I call ‘post processing’ networks that take output from the camera networks and turn it into higher level actionable abstractions. So far I’ve only looked at the camera networks for V9 but it’s already apparent that V9 is a pretty big change from V8.

    ---------------
    One unified camera network handles all 8 cameras

    Same weight file being used for all cameras (this has pretty interesting implications and previously V8 main/narrow seems to have had separate weights for each camera)

    Processed resolution of 3 front cameras and back camera: 1280x960 (full camera resolution)

    Processed resolution of pillar and repeater cameras: 640x480 (1/2x1/2 of camera’s true resolution)

    all cameras: 3 color channels, 2 frames (2 frames also has very interesting implications)

    (was 640x416, 2 color channels, 1 frame, only main and narrow in V8)
    ------------

    Various V8 versions included networks for pillar and repeater cameras in the binaries but AFAIK nobody outside Tesla ever saw those networks in operation. Normal AP use on V8 seemed to only include the use of main and narrow for driving and the wide angle forward camera for rain sensing. In V9 it’s very clear that all cameras are being put to use for all the AP2 cars.

    The basic camera NN (neural network) arrangement is an Inception V1 type CNN with L1/L2/L3ab/L4abcdefg layer arrangement (architecturally similar to V8 main/narrow camera up to end of inception blocks but much larger)
    • about 5x as many weights as comparable portion of V8 net
    • about 18x as much processing per camera (front/back)
    The V9 network takes 1280x960 images with 3 color channels and 2 frames per camera from, for example, the main camera. That’s 1280x960x3x2 as an input, or 7.3M. The V8 main camera was 640x416x2 or 0.5M - 13x less data.

    For perspective, V9 camera network is 10x larger and requires 200x more computation when compared to Google’s Inception V1 network from which V9 gets it’s underlying architectural concept. That’s processing *per camera* for the 4 front and back cameras. Side cameras are 1/4 the processing due to being 1/4 as many total pixels. With all 8 cameras being processed in this fashion it’s likely that V9 is straining the compute capability of the APE. The V8 network, by comparison, probably had lots of margin.

    network outputs:
    • V360 object decoder (multi level, processed only)
    • back lane decoder (back camera plus final processed)
    • side lane decoder (pillar/repeater cameras plus final processed)
    • path prediction pp decoder (main/narrow/fisheye cameras plus final processed)
    • “super lane” decoder (main/narrow/fisheye cameras plus final processed)

    Previous V8 aknet included a lot of processing after the inception blocks - about half of the camera network processing was taken up by non-inception weights. V9 only includes inception components in the camera network and instead passes the inception processed outputs, raw camera frames, and lots of intermediate results to the post processing subsystem. I have not yet examined the post processing subsystem.

    And now for some speculation:

    Input changes:

    The V9 network takes 1280x960 images with 3 color channels and 2 frames per camera from, for example, the main camera. That’s 1280x960x3x2 as an input, or 7.3MB. The V8 main camera processing frame was 640x416x2 or 0.5MB - 13x less data. The extra resolution means that V9 has access to smaller and more subtle detail from the camera, but the more interesting aspect of the change to the camera interface is that camera frames are being processed in pairs. These two pairs are likely time-offset by some small delay - 10ms to 100ms I’d guess - allowing each processed camera input to see motion. Motion can give you depth, separate objects from the background, help identify objects, predict object trajectories, and provide information about the vehicle’s own motion. It's a pretty fundamental improvement to the basic perceptions of the system.

    Camera agnostic:

    The V8 main/narrow network used the same architecture for both cameras, but by my calculation it was probably using different weights for each camera (probably 26M each for a total of about 52M). This make sense because main/narrow have a very different FOV, which means the precise shape of objects they see varies quite a bit - especially towards the edges of frames. Training each camera separately is going to dramatically simplify the problem of recognizing objects since the variation goes down a lot. That means it’s easier to get decent performance with a smaller network and less training. But it also means you have to build separate training data sets, evaluate them separately, and load two different networks alternately during operation. It also means that you network can learn some bad habits because it always sees the world in the same way.

    Building a camera agnostic network relaxes these problems and simultaneously makes the network more robust when used on any individual camera. Being camera agnostic means the network has to have a better sense of what an object looks like under all kinds of camera distortions. That’s a great thing, but it’s very, *very* expensive to achieve because it requires a lot of training, a lot of training data and, probably, a really big network. Nobody builds them so it’s hard to say for sure, but these are probably safe assumptions.

    Well, the V9 network appears to be camera agnostic. It can process the output from any camera on the car using the same weight file.

    It also has the fringe benefit of improved computational efficiency. Since you just have the one set of weights you don’t have to constantly be swapping weight sets in and out of your GPU memory and, even more importantly, you can batch up blocks of images from all the cameras together and run them through the NN as a set. This can give you a multiple of performance from the same hardware.

    I didn’t expect to see a camera agnostic network for a long time. It’s kind of shocking.

    Considering network size:

    This V9 network is a monster, and that’s not the half of it. When you increase the number of parameters (weights) in an NN by a factor of 5 you don’t just get 5 times the capacity and need 5 times as much training data. In terms of expressive capacity increase it’s more akin to a number with 5 times as many digits. So if V8’s expressive capacity was 10, V9’s capacity is more like 100,000. It’s a mind boggling expansion of raw capacity. And likewise the amount of training data doesn’t go up by a mere 5x. It probably takes at least thousands and perhaps millions of times more data to fully utilize a network that has 5x as many parameters.

    This network is far larger than any vision NN I’ve seen publicly disclosed and I’m just reeling at the thought of how much data it must take to train it. I sat on this estimate for a long time because I thought that I must have made a mistake. But going over it again and again I find that it’s not my calculations that were off, it’s my expectations that were off.

    Is Tesla using semi-supervised training for V9? They've gotta be using more than just labeled data - there aren't enough humans to label this much data. I think all those simulation designers they hired must have built a machine that generates labeled data for them, but even so.

    And where are they getting the datacenter to train this thing? Did Larry give Elon a warehouse full of TPUs?

    I mean, seriously...

    I look at this thing and I think - oh yeah, HW3. We’re gonna need that. Soon, I think.

    Omnidirectionality (V360 object decoder):

    With these new changes the NN should be able to identify every object in every direction at distances up to hundreds of meters and also provide approximate instantaneous relative movement for all of those objects. If you consider the FOV overlap of the cameras, virtually all objects will be seen by at least two cameras. That provides the opportunity for downstream processing use multiple perspectives on an object to more precisely localize and identify it.

    General thoughts:

    I’ve been driving V9 AP2 for a few days now and I find the dynamics to be much improved over recent V8. Lateral control is tighter and it’s been able to beat all the V8 failure scenarios I’ve collected over the last 6 months. Longitudinal control is much smoother, traffic handling is much more comfortable. V9’s ability to prospectively do a visual evaluation on a target lane prior to making a change makes the auto lane change feature a lot more versatile. I suspect detection errors are way down compared to V8 but I also see that a few new failure scenarios have popped up (offramp / onramp speed control seem to have some bugs). I’m excited to see how this looks in a couple of months after they’ve cleaned out the kinks that come with any big change.

    Being an avid observer of progress in deep neural networks my primary motivation for looking at AP2 is that it’s one of the few bleeding edge commercial applications that I can get my hands on and I use it as a barometer of how commercial (as opposed to research) applications are progressing. Researchers push the boundaries in search of new knowledge, but commercial applications explore the practical ramifications of new techniques. Given rapid progress in algorithms I had expected near future applications might hinge on the great leaps in efficiency that are coming from new techniques. But that’s not what seems to be happening right now - probably because companies can do a lot just by scaling up NN techniques we already have.

    In V9 we see Tesla pushing in this direction. Inception V1 is a four year old architecture that Tesla is scaling to a degree that I imagine inceptions’s creators could not have expected. Indeed, I would guess that four years ago most people in the field would not have expected that scaling would work this well. Scaling computational power, training data, and industrial resources plays to Tesla’s strengths and involves less uncertainty than potentially more powerful but less mature techniques. At the same time Tesla is doubling down on their ‘vision first / all neural networks’ approach and, as far as I can tell, it seems to be going well.

    As a neural network dork I couldn’t be more pleased.
     
    • Informative x 165
    • Love x 66
    • Like x 21
    • Helpful x 7
  10. pkodali

    pkodali Banned

    Joined:
    Oct 22, 2017
    Messages:
    817
    Location:
    Sacramento
    I didn't understand half of what you said but thank you so much for taking the time to write this
     
    • Like x 30
    • Funny x 10
    • Love x 5
    • Informative x 1
  11. DeckardsGirl

    DeckardsGirl Deckard my Model 3 owns me

    Joined:
    Jan 20, 2018
    Messages:
    26
    Location:
    Austin
    Thank you jimmie_d - great analysis for everyone on the neural network behind V9. As always you peel back the cover so we can all understand!
     
    • Like x 5
    • Love x 1
  12. Alketi

    Alketi Member

    Joined:
    May 3, 2016
    Messages:
    851
    Location:
    Boston, North Shore
    @jimmy_d I LOVE your NN breakdowns. Please don't ever stop. :)

    A couple of observations on the NN behavior that I'd love to have your perspective on.

    1. The NN seems to improperly identify stationary objects as buses or trucks -- i.e. cars parked close together, various objects in one's driveway, etc. These objects flip into existence and then sometimes jump between different object types as they're constantly re-identified. Do you think that the NN has a deliberately low threshold applied to its confidence setting, such that these objects can and will be mis-identified? That perhaps its safer at this stage to misidentify a real thing, rather than fail to classify it properly and potentially ignore it?

    2. Vehicles in adjacent lanes seem to have a high degree of "float". The appear on screen to cross lane lines, etc. Yet the car is not reacting to these misplacements. So, there would seem to be a de-coupling of the renderings vs. the behavior. Do you think this implies that the car is not yet trusting their locations?

    My thought behind #2 is that even with v9 the driving network still has't fully mastered the merging traffic scenario, where it will smoothly anticipate a vehicle coming into your lane. And to master that, I imagine it must rely on 3D positioning from the cameras, as the radar is perhaps too forward looking. Once this is solved, I think user disengagement events will drop dramatically as this one of the last problem areas at least for straight driving.

    Thanks again.
     
    • Like x 9
    • Helpful x 3
    • Informative x 1
    • Love x 1
  13. jimmy_d

    jimmy_d Deep Learning Dork

    Joined:
    Jan 26, 2016
    Messages:
    412
    Location:
    San Francisco, CA
    That's a great sentiment. Thank you.
     
    • Like x 8
    • Love x 3
    • Informative x 1
  14. jimmy_d

    jimmy_d Deep Learning Dork

    Joined:
    Jan 26, 2016
    Messages:
    412
    Location:
    San Francisco, CA
    Was having a conversation about this with someone the other day. I think the display behavior you point out must be confusing to a lot of people because I've heard it from pretty much everybody. It's obvious when you think about it, but what appears on the display is not what the car is using to drive, it's a separate output that is just there to give the driver some utility. If the car's driving decision's were based on the same interpretation of that display data that a human uses it would be braking and swerving all over the place. But the car doesn't do that. The display info is just for driver consumption and it's obviously still got some problems.

    The display stuff is probably getting interpreted heuristically from the camera network outputs. The camera networks provide a smooth probabilistic output that is not very easy to display with a human written algorithm. But you can't put the NN output directly on the display either. So the display code probably looks at periodic snapshots of the camera network outputs in order to provide a low latency output and it has to have a hard threshold for deciding whether some vehicle 'exists' at some point in space because, seriously, a cloud of vehicle expectation values distributed over 3-space is not something you can put on a car screen.

    But it's not a problem for the car driving itself because the driving decision networks get the full feed and get to watch all those probabilities vary smoothly over time. Those networks have a much more nuanced interpretation of what's coming out of the camera networks than the display code does.

    Ideally that display should be a reliable window into the perceptions of the car. It's not there yet.
     
    • Informative x 60
    • Like x 10
    • Helpful x 6
    • Disagree x 1
  15. generalenthu

    generalenthu Member

    Joined:
    Jun 10, 2015
    Messages:
    740
    Location:
    Vienna, VA
    Thanks a ton @jimmy_d for all your sleuthing. Couple of questions.

    Any thoughts on what Elon is saying here about different NNs for various cameras? Was it just misdirection?

    Twitter

    Secondly, there were some tweets by Karpathy a few months ago about distributed training of NNs. With a quickly increasing fleet of this size, most of it plugged in and on Wifi at night, is it possible they are or can tap into this idle compute?
     
    • Informative x 2
    • Like x 1
  16. strangecosmos

    strangecosmos Non-Member

    Joined:
    May 10, 2017
    Messages:
    1,041
    Location:
    The Prime Material Plane
    #516 strangecosmos, Oct 14, 2018
    Last edited: Oct 14, 2018
    Do you mean 2 frames per second from each camera? Is that all? And was it only 1 frame per second until now?? :eek:

    P.S. Thanks for your incredible analysis. Makes TMC worth coming back to.
     
    • Like x 2
    • Funny x 1
  17. animorph

    animorph Active Member

    Joined:
    Apr 1, 2016
    Messages:
    1,448
    Location:
    Scottsdale, AZ
    No. He meant it processes 2 frames at a time so that it should be able to see movement. Not an indication of frame rate.
     
    • Helpful x 6
    • Informative x 2
    • Like x 2
  18. strangecosmos

    strangecosmos Non-Member

    Joined:
    May 10, 2017
    Messages:
    1,041
    Location:
    The Prime Material Plane
    #518 strangecosmos, Oct 14, 2018
    Last edited: Oct 14, 2018
    Ahhh, gotcha, thanks. Well, that is very cool.

    How much training data do you think is needed?

    You can rent a TPU v3 from Google for $2.40/hour. AWS does GPUs for $0.05/hour. Given Tesla spent $386 million on R&D in Q2, I bet the AI team has the budget for quite a bit of compute.

    A budget of $25 million/quarter would be enough for 10.4 million TPU v3 hours/quarter or 500 million GPU hours/quarter. How much training could that accomplish?

    I’m curious to know what the newer techniques are!
     
  19. gluu

    gluu Member

    Joined:
    Jan 18, 2018
    Messages:
    72
    Location:
    Chicago
    So with one network shared on all cameras, I think this likely means that Tesla will be less sensitive to hardware differences and changes across vehicles. I read somewhere recently Cruise had a lot of trouble with the new Bolt architecture because sensor locations were changed.
     
    • Informative x 3
    • Like x 1
  20. mhan00

    mhan00 Member

    Joined:
    Oct 13, 2014
    Messages:
    737
    Location:
    Southern California
    Me reading your post: :eek::confused::D

    Great breakdown and super informative, just not to me since I’m too dumb to understand like 90 percent of what you were talking about. It sounded awesome and encouraging though. Looking forward to more insights from you; hopefully over time some of your knowledge will begin to sink in, lol.
     
    • Like x 1
    • Funny x 1

Share This Page

  • About Us

    Formed in 2006, Tesla Motors Club (TMC) was the first independent online Tesla community. Today it remains the largest and most dynamic community of Tesla enthusiasts. Learn more.
  • Do you value your experience at TMC? Consider becoming a Supporting Member of Tesla Motors Club. As a thank you for your contribution, you'll get nearly no ads in the Community and Groups sections. Additional perks are available depending on the level of contribution. Please visit the Account Upgrades page for more details.


    SUPPORT TMC