3 ms·

The big leap in deep neural nets, gave us the illusion that we will solve many of our problems overnight. The fact is that we probably need at least one more de

The big leap in deep neural nets, gave us the illusion that we will solve many of our problems overnight. The fact is that we probably need at least one more decade to get things right

vokep 9y ago
The leap was that now we have a plan for how to tackle these problems, we have tools that will actually work. Seemingly overnight those tools appeared and we realized their potential. However, we still need to learn how to best use these new tools and perfect them. Just because we can do something doesn't mean we can do it good enough, yet.
- pzone 9y ago
  But it's still easy to sucker 3 grand from all the people who didn't realize this.
- nopriorarrests 9y ago
  Do we know the tools that will work? Waymo uses expensive LIDAR's, and we dont know the price of the whole hardware set, since Waymo does not care about the price, they work on scientific problem. Tesla claims that cameras are enough and they can put it into mass-produced cars without changing the cost significantly. Pretty large gap, don't you think?
- ghaff 9y ago
  And you have to assume that optimizing this particular toolset gets you 100% (or at least approaching 100% in the limit) to solving the problem that you need to solve. Deep learning/neural nets popped for a variety of factors relative to a number of other techniques that had also been hanging around for ages. But there's no guarantee that past performance will be predictive of future results.
imtringued 9y ago
Yes definitively but if Waymo is a decade away then how far away from truly self driving cars is Tesla? 15 years? 20 years? Tesla claims that LIDAR isn't neccessary but yet they constantly hit barriers and stopped trucks. If they can't even get the basics right how can Musk be sure that his strategy is the right one?
- thebluehawk 9y ago
  I don't disagree with your points, other than my definition of "constantly" must be different than yours.
- buvanshak 9y ago
  > If they can't even get the basics right how can Musk be sure that his strategy is the right one? The strategy is right one as long as it sells cars. Musk is a salesman in engineer's clothing.
Animats 9y ago
Neural nets are only a small part of the solution. Waymo gets this, but too many of the others overuse neural nets. The safe solutions are about 80% geometry and 20% neural net. LIDAR and geometry give you a 3D map of the environment, at least the part of it in range of the sensors. Anywhere the vehicle is going to drive should be verified as flat. That's the first step. Waymo does this. Tesla doesn't. Whether Uber does will come out when the NTSB dissects their vehicle that ran over a pedestrian. Those top-mounted LIDAR units aren't primarily looking for obstacles. They're looking for flat road. A pothole, a cliff, a car, or a pedestrian does not look like flat road. The software doesn't have to know why it's not flat to not go there. Trying to classify stuff that isn't flat road is mostly for dealing with moving objects. "What is that?" and "What will it do next" isn't a geometry problem. That's where neural nets come in. An acceptable result is "unidentified", which feeds back into the planner as "stay away from there". Some people have gotten the idea that they just hook up a camera to a machine learning system, train it from a dashcam, and they have a quick path to self-driving. That will work most of the time. Until it kills someone.

The big leap in deep neural nets, gave us the illusion that we will solve many of our problems overnight. The fact is that we probably need at least one more decade to get things right

vokep 9y ago

The leap was that now we have a plan for how to tackle these problems, we have tools that will actually work. Seemingly overnight those tools appeared and we realized their potential. However, we still need to learn how to best use these new tools and perfect them. Just because we can do something doesn't mean we can do it good enough, yet.

pzone 9y ago
But it's still easy to sucker 3 grand from all the people who didn't realize this.
nopriorarrests 9y ago
Do we know the tools that will work? Waymo uses expensive LIDAR's, and we dont know the price of the whole hardware set, since Waymo does not care about the price, they work on scientific problem. Tesla claims that cameras are enough and they can put it into mass-produced cars without changing the cost significantly. Pretty large gap, don't you think?
ghaff 9y ago
And you have to assume that optimizing this particular toolset gets you 100% (or at least approaching 100% in the limit) to solving the problem that you need to solve. Deep learning/neural nets popped for a variety of factors relative to a number of other techniques that had also been hanging around for ages. But there's no guarantee that past performance will be predictive of future results.

imtringued 9y ago

Yes definitively but if Waymo is a decade away then how far away from truly self driving cars is Tesla? 15 years? 20 years? Tesla claims that LIDAR isn't neccessary but yet they constantly hit barriers and stopped trucks. If they can't even get the basics right how can Musk be sure that his strategy is the right one?

thebluehawk 9y ago
I don't disagree with your points, other than my definition of "constantly" must be different than yours.
buvanshak 9y ago
> If they can't even get the basics right how can Musk be sure that his strategy is the right one? The strategy is right one as long as it sells cars. Musk is a salesman in engineer's clothing.

Animats 9y ago

Neural nets are only a small part of the solution. Waymo gets this, but too many of the others overuse neural nets. The safe solutions are about 80% geometry and 20% neural net. LIDAR and geometry give you a 3D map of the environment, at least the part of it in range of the sensors. Anywhere the vehicle is going to drive should be verified as flat. That's the first step. Waymo does this. Tesla doesn't. Whether Uber does will come out when the NTSB dissects their vehicle that ran over a pedestrian. Those top-mounted LIDAR units aren't primarily looking for obstacles. They're looking for flat road. A pothole, a cliff, a car, or a pedestrian does not look like flat road. The software doesn't have to know why it's not flat to not go there. Trying to classify stuff that isn't flat road is mostly for dealing with moving objects. "What is that?" and "What will it do next" isn't a geometry problem. That's where neural nets come in. An acceptable result is "unidentified", which feeds back into the planner as "stay away from there". Some people have gotten the idea that they just hook up a camera to a machine learning system, train it from a dashcam, and they have a quick path to self-driving. That will work most of the time. Until it kills someone.