3 ms·

The table goes into more detail. Each time the classification changed, the history for that object was essentially deleted; since there's only one data point, t

The table goes into more detail. Each time the classification changed, the history for that object was essentially deleted; since there's only one data point, the system predicted that it would "continue" to stay stationary, even though the pedestrian was walking at a steady pace.

rubicon33 7y ago
That part, about deleting the history, confuses me. Why delete the history on a classification change? Shouldn't classifications be tiered? In this case, while the system was struggling to PERFECTLY classify the object, it was clearly thinking it was something that should be avoided (oscillating between car, bike, other). In this case, I would expect the system to keep it's motion history. IMO, this could have prevented the accident because although it didn't determine it was a bicycle/person until "too late" ... it had determined with plenty of time that it was maybe a car, maybe a bike.
- beerandt 7y ago
  If it's having a hard time both identifying an object, as well as measuring it's movement, there's not really any reason it should understand that all those separate data points are the same object. That is, it doesn't really matter if the object history is "deleted" or not; if it can't associate a new data point with a previous history (by identification or predicted position), the practical result is the same as if there is no object history. This could be a result of using velocity based tracking, which I don't know that Uber uses, but is a fairly standard method, as it's what raw GPS measurements are based on.
  - Udik 7y ago
    > if it can't associate a new data point with a previous history (by identification or predicted position) This sounds a bit strange, given that the entire technology around which autonomous driving is built is about identifying and recognising patterns in extremely noisy and variable data.
- SilasX 7y ago
  This classification failure reminds me of the concept of "Disguised queries" I saw on LessWrong [1]. That is, there's the question you really want to answer: A) Should I avoid driving this car into this thing? and then there are subsidiary questions that help to answer that, but would be fully obviated if you already had a good answer for A): B1) Is this a human? B2) Is this a vehicle? B3) Is this unrecognizable? The "natural category" here is "drive into" vs "don't drive into", not "human vs vehicle vs fruit stand vs other". [1] https://www.lesswrong.com/posts/4FcxgdvdQP45D6Skg/disguised-queries https://www.lesswrong.com/posts/4FcxgdvdQP45D6Skg/disguised-... In the article, the A)-type question is whether the object has vanadium, and the B)-type questions are whether the object is a blegg (blue round thing) or a rube (red square thing). The distinction becomes stark when you know it has vanadium, but it doesn't neatly fall on the blegg/rube continuum.

rubicon33 7y ago

That part, about deleting the history, confuses me. Why delete the history on a classification change? Shouldn't classifications be tiered? In this case, while the system was struggling to PERFECTLY classify the object, it was clearly thinking it was something that should be avoided (oscillating between car, bike, other). In this case, I would expect the system to keep it's motion history. IMO, this could have prevented the accident because although it didn't determine it was a bicycle/person until "too late" ... it had determined with plenty of time that it was maybe a car, maybe a bike.

beerandt 7y ago
If it's having a hard time both identifying an object, as well as measuring it's movement, there's not really any reason it should understand that all those separate data points are the same object. That is, it doesn't really matter if the object history is "deleted" or not; if it can't associate a new data point with a previous history (by identification or predicted position), the practical result is the same as if there is no object history. This could be a result of using velocity based tracking, which I don't know that Uber uses, but is a fairly standard method, as it's what raw GPS measurements are based on.
- Udik 7y ago
  > if it can't associate a new data point with a previous history (by identification or predicted position) This sounds a bit strange, given that the entire technology around which autonomous driving is built is about identifying and recognising patterns in extremely noisy and variable data.
SilasX 7y ago
This classification failure reminds me of the concept of "Disguised queries" I saw on LessWrong [1]. That is, there's the question you really want to answer: A) Should I avoid driving this car into this thing? and then there are subsidiary questions that help to answer that, but would be fully obviated if you already had a good answer for A): B1) Is this a human? B2) Is this a vehicle? B3) Is this unrecognizable? The "natural category" here is "drive into" vs "don't drive into", not "human vs vehicle vs fruit stand vs other". [1] https://www.lesswrong.com/posts/4FcxgdvdQP45D6Skg/disguised-queries https://www.lesswrong.com/posts/4FcxgdvdQP45D6Skg/disguised-... In the article, the A)-type question is whether the object has vanadium, and the B)-type questions are whether the object is a blegg (blue round thing) or a rube (red square thing). The distinction becomes stark when you know it has vanadium, but it doesn't neatly fall on the blegg/rube continuum.