4 ms·

I think you might not be familiar with the concept of defensive driving. Defensive driving means slowing down and paying more attention in areas of low visibil

I think you might not be familiar with the concept of defensive driving. Defensive driving means slowing down and paying more attention in areas of low visibility, especially if there are likely pedestrian crossings. It means increasing your following distance. It means increasing the odds that even if something unlikely happens you, the driver, will still be in a better position to react, or you will be less likely to cause harm. It's not about prevention, it's about risk management and harm reduction. This is definitely something that is likely to be difficult to teach to autonomous systems.

danmaz74 9y ago
> This is definitely something that is likely to be difficult to teach to autonomous systems. Why? What you listed seem like conditions that should be easy enough to recognize even with simple algorithms.
- TeMPOraL 9y ago
  Yup. An algorithm should be in much better position to recognize its confidence about what it sees and can happen, and alter the driving speed to reduce the amount of uncertain spots that fall within its stopping range. That is, unless they use neural network magic black boxes to implement the self-driving part.
  - carlmr 9y ago
    >That is, unless they use neural network magic black boxes to implement the self-driving part. Is this actually done? I thought neural networks were more or less used for the vision system to detect parts of the scene (pedestrians, stop-signs, etc). This is the only area where to me it makes sense to take the blackbox tadeoff right now.
    - rcxdude 9y ago
      Neural nets are also likely to be used to model the behavior of other road users (i.e. 'how is this car/pedestrian/etc likely to move in the next few seconds'?). No-one seems to be seriously considering an end-to-end neural network (i.e. inputs are sensors, outputs are throttle and steering wheel), but neural nets are pervasive components of the full system.

danmaz74 9y ago

> This is definitely something that is likely to be difficult to teach to autonomous systems. Why? What you listed seem like conditions that should be easy enough to recognize even with simple algorithms.

TeMPOraL 9y ago
Yup. An algorithm should be in much better position to recognize its confidence about what it sees and can happen, and alter the driving speed to reduce the amount of uncertain spots that fall within its stopping range. That is, unless they use neural network magic black boxes to implement the self-driving part.
- carlmr 9y ago
  >That is, unless they use neural network magic black boxes to implement the self-driving part. Is this actually done? I thought neural networks were more or less used for the vision system to detect parts of the scene (pedestrians, stop-signs, etc). This is the only area where to me it makes sense to take the blackbox tadeoff right now.
  - rcxdude 9y ago
    Neural nets are also likely to be used to model the behavior of other road users (i.e. 'how is this car/pedestrian/etc likely to move in the next few seconds'?). No-one seems to be seriously considering an end-to-end neural network (i.e. inputs are sensors, outputs are throttle and steering wheel), but neural nets are pervasive components of the full system.