3 ms·

> The biggest difference with previous FSD updates is that the vehicle’s controls would now be handled by neural nets rather than being coded by programmers. D

> The biggest difference with previous FSD updates is that the vehicle’s controls would now be handled by neural nets rather than being coded by programmers. Does anyone have details on what this means? To me this sounds like neural nets are now part of the low level control loops for the throttle/steering as well - but that doesn't seem particularly beneficial. Especially if you're interested in control stability/predictability.

HarHarVeryFunny 3y ago
Yeah - Tesla software has two parts, a vision system and a driving system. The vision system turns the camera image into a description of what is in the image (road, cars, people, etc), and the driving system actually controls the car based on the vision input and route it has been given. Up until now the driving part had been hand coded in C++, but now it is a neural net that has learnt what would a human do in similar road situation (not sure what they trained it on). Is this an improvement? Who knows - should certainly be different. This doesn't mean the car suddenly got smart and now understands what it is looking at - it just means it's using a different set of rules for when to switch lanes, brake, etc.
- cdumler 3y ago
  I would argue reasonable to believe it will be notably better until it is catastrophically wrong. The issue with neural nets is that they can be really well trained within the bounds of training set; however, when faced with novel, there is no underlying abstraction inference on which to revert. Look instances where autonomous cars are disable or have weird responses when people modify signs[1]. The driving will be better, but the when the information sits squarely outside of the norm (art on the road, abstract images that defy normal depth cues, etc), the car will do very unexpected things. As humans, we have this issue already. We call them "optical illusions." Optical illusions are really brain neural net failures where there are competing narratives between evaluations of the information provided by our senses we conclude invalid results. We don't have the ability to process absolute light values, so we infer colorspace by relative color within our 3D mental model, which can result in inferring a difference in color where there is none. 1. https://spectrum.ieee.org/slight-street-sign-modifications-can-fool-machine-learning-algorithms https://spectrum.ieee.org/slight-street-sign-modifications-c... 2. https://en.wikipedia.org/wiki/Checker_shadow_illusion https://en.wikipedia.org/wiki/Checker_shadow_illusion
andrewmcwatters 3y ago
You’re right. comma.ai is making the same mistake and it shows in their current experimental mode. I find the argument that you can allow both longitudinal and lateral control to be expressed by a black box difficult to accept. The proof of the driving is in the handling. There’s a video of an engineer at comma.ai mocking self-driving teams for semantically labeling objects on the road in SAE Level 2 and beyond systems. Which screams stupid to me. Why would you be proud that your system has no idea what a pedestrian is. Or what a traffic cone or a speed limit sign are? Idiotic. At least Tesla still semantically labels. But I argue that you can’t get smooth driving out of neural network controls because they inherently have no idea what smooth driving explicitly is. They estimate it. The best they can do is attempt to mock it, which without programmer-defined rails WILL result in excessively high jerk and acceleration numbers, as proven by openpilot’s implementation.
- zaptrem 3y ago
  The problem is manual labelling doesn't scale. Switching to an unsupervised approach is what made many huge AI advances over the past couple years possible since we're mostly limited by compute now. For Tesla (who have practically unlimited driving footage) this is particularly advantageous.
  - andrewmcwatters 3y ago
    It’s not the labeling itself, it’s whether or not you’re allowing unfiltered longitudinal control to originate from neural nets. Good but erratic lateral control at worst results in hugging a lane line. Anything less than perfect longitudinal control means the driver will toggle it on and off during portions of traffic where it won’t follow the flow of traffic closely enough, or if it follows to closely, or if you’re experiencing a high cabin jerk factor. You cannot screw up longitudinal control. It’s nearly the whole experience.
    - zaptrem 3y ago
      I think this is what end-to-end models will be best at, since they'll be able to account for far more environmental factors than a manually written planner flipping between the stop light deceleration curve and car-ahead curve constantly, for example.
hulitu 3y ago
Yes it means there is no difference. Neural nets are coded by programmers.

HarHarVeryFunny 3y ago

Yeah - Tesla software has two parts, a vision system and a driving system. The vision system turns the camera image into a description of what is in the image (road, cars, people, etc), and the driving system actually controls the car based on the vision input and route it has been given. Up until now the driving part had been hand coded in C++, but now it is a neural net that has learnt what would a human do in similar road situation (not sure what they trained it on). Is this an improvement? Who knows - should certainly be different. This doesn't mean the car suddenly got smart and now understands what it is looking at - it just means it's using a different set of rules for when to switch lanes, brake, etc.

cdumler 3y ago
I would argue reasonable to believe it will be notably better until it is catastrophically wrong. The issue with neural nets is that they can be really well trained within the bounds of training set; however, when faced with novel, there is no underlying abstraction inference on which to revert. Look instances where autonomous cars are disable or have weird responses when people modify signs[1]. The driving will be better, but the when the information sits squarely outside of the norm (art on the road, abstract images that defy normal depth cues, etc), the car will do very unexpected things. As humans, we have this issue already. We call them "optical illusions." Optical illusions are really brain neural net failures where there are competing narratives between evaluations of the information provided by our senses we conclude invalid results. We don't have the ability to process absolute light values, so we infer colorspace by relative color within our 3D mental model, which can result in inferring a difference in color where there is none. 1. https://spectrum.ieee.org/slight-street-sign-modifications-can-fool-machine-learning-algorithms https://spectrum.ieee.org/slight-street-sign-modifications-c... 2. https://en.wikipedia.org/wiki/Checker_shadow_illusion https://en.wikipedia.org/wiki/Checker_shadow_illusion

andrewmcwatters 3y ago

You’re right. comma.ai is making the same mistake and it shows in their current experimental mode. I find the argument that you can allow both longitudinal and lateral control to be expressed by a black box difficult to accept. The proof of the driving is in the handling. There’s a video of an engineer at comma.ai mocking self-driving teams for semantically labeling objects on the road in SAE Level 2 and beyond systems. Which screams stupid to me. Why would you be proud that your system has no idea what a pedestrian is. Or what a traffic cone or a speed limit sign are? Idiotic. At least Tesla still semantically labels. But I argue that you can’t get smooth driving out of neural network controls because they inherently have no idea what smooth driving explicitly is. They estimate it. The best they can do is attempt to mock it, which without programmer-defined rails WILL result in excessively high jerk and acceleration numbers, as proven by openpilot’s implementation.

zaptrem 3y ago
The problem is manual labelling doesn't scale. Switching to an unsupervised approach is what made many huge AI advances over the past couple years possible since we're mostly limited by compute now. For Tesla (who have practically unlimited driving footage) this is particularly advantageous.
- andrewmcwatters 3y ago
  It’s not the labeling itself, it’s whether or not you’re allowing unfiltered longitudinal control to originate from neural nets. Good but erratic lateral control at worst results in hugging a lane line. Anything less than perfect longitudinal control means the driver will toggle it on and off during portions of traffic where it won’t follow the flow of traffic closely enough, or if it follows to closely, or if you’re experiencing a high cabin jerk factor. You cannot screw up longitudinal control. It’s nearly the whole experience.
  - zaptrem 3y ago
    I think this is what end-to-end models will be best at, since they'll be able to account for far more environmental factors than a manually written planner flipping between the stop light deceleration curve and car-ahead curve constantly, for example.

hulitu 3y ago

Yes it means there is no difference. Neural nets are coded by programmers.