4 ms·

Why would machines be worse than humans in bad weather if they are better than humans in good weather? Its not like our eyes are better than the machines camera

Why would machines be worse than humans in bad weather if they are better than humans in good weather? Its not like our eyes are better than the machines cameras. If its because human drivers can better "build a mental model" of whats happening with limited visibility/data - that seems like the same problem that autonomous cars are designed to solve. Whats stopping autonomous cars from making progress in the bad weather arena? Is it just we haven't done enough miles in that weather and we need to train the ML some more?

teraflop 11y ago
You're putting a lot of faith in the ability to solve a very difficult problem. Computer vision has made huge strides in recent years, but it's still far behind the capabilities of the human visual cortex. It's not just a matter of building a simple model and training it until it achieves arbitrarily high accuracy. ML encompasses a powerful set of tools, but it's not a magic black box that eats data and spits out AI. We don't yet know how the human brain processes images at a detailed level, and it's too complex of a system to optimize by brute-force. The best ML image processing systems still rely on a lot of pre-encoded assumptions, and require huge amounts of training data/computational resources, and don't perform as well as humans even on simple tasks. For now, autonomous vehicles "cheat" a lot by using sensor technologies like lidar and radar -- techniques that produce spatial data that can be interpreted more easily than raw imagery. That adds a lot to the hardware costs, but it's currently the only way we can make them perform acceptably well in real-world scenarios. And conditions that interfere with those technologies more severely, like fog or snow, are still crippling.
- JoeAltmaier 11y ago
  Actually there's no reason to suppose that computer vision hasn't already exceeded the human capacity. After all, it has lidar and radar. Cheating or not, it has a very good result.
  - teraflop 11y ago
    Agreed -- don't knock it if it works! However, those kinds of sensors don't really fall into what's normally considered "computer vision". Mind you, I'm extremely optimistic about the potential successes of autonomous cars in the near- to mid-future (the next 5 to 25 years, say). But I don't want people to get the wrong impression about where the state of the art is today. And in particular, I want to push back against the idea that machine learning is a solved problem, and that autonomous cars can teach themselves to drive as well as a human if we only give them enough miles of practice.
rtkwe 11y ago
Vision algorithms are pretty shaky to begin with. We've made leaps and bounds of improvements but they're still pretty far behind what people are able to do. In inclement conditions I think humans still out perform machines pretty handily. Eventually it'll be an even field or there will be sensors added to cars that can just see through the snow/rain but that's not where we are.
TwiztidK 11y ago
> Why would machines be worse than humans in bad weather if they are better than humans in good weather? [Because they are.](http://jalopnik.com/this-is-how-bad-self-driving-cars-suck-in-the-rain-1666268433 http://jalopnik.com/this-is-how-bad-self-driving-cars-suck-i...) Basically, the sensors they tend to equipped with aren't even close to being sufficient in poor whether. Humans are EXTREMELY good at pattern recognition and can manage to make it work, but computers just aren't at that level yet. It'll take a combination of improved sensors and computer vision (ie, pattern recognition) to make it happen.
greedo 11y ago
Because Google has mapped out the roads to a very specific level; this allows them to use sensors in the cars that don't have to have the fidelity of human eyes etc. As Google's maps become out of date, they need to be recreated or checked. Their system won't scale...

teraflop 11y ago

You're putting a lot of faith in the ability to solve a very difficult problem. Computer vision has made huge strides in recent years, but it's still far behind the capabilities of the human visual cortex. It's not just a matter of building a simple model and training it until it achieves arbitrarily high accuracy. ML encompasses a powerful set of tools, but it's not a magic black box that eats data and spits out AI. We don't yet know how the human brain processes images at a detailed level, and it's too complex of a system to optimize by brute-force. The best ML image processing systems still rely on a lot of pre-encoded assumptions, and require huge amounts of training data/computational resources, and don't perform as well as humans even on simple tasks. For now, autonomous vehicles "cheat" a lot by using sensor technologies like lidar and radar -- techniques that produce spatial data that can be interpreted more easily than raw imagery. That adds a lot to the hardware costs, but it's currently the only way we can make them perform acceptably well in real-world scenarios. And conditions that interfere with those technologies more severely, like fog or snow, are still crippling.

JoeAltmaier 11y ago
Actually there's no reason to suppose that computer vision hasn't already exceeded the human capacity. After all, it has lidar and radar. Cheating or not, it has a very good result.
- teraflop 11y ago
  Agreed -- don't knock it if it works! However, those kinds of sensors don't really fall into what's normally considered "computer vision". Mind you, I'm extremely optimistic about the potential successes of autonomous cars in the near- to mid-future (the next 5 to 25 years, say). But I don't want people to get the wrong impression about where the state of the art is today. And in particular, I want to push back against the idea that machine learning is a solved problem, and that autonomous cars can teach themselves to drive as well as a human if we only give them enough miles of practice.

rtkwe 11y ago

Vision algorithms are pretty shaky to begin with. We've made leaps and bounds of improvements but they're still pretty far behind what people are able to do. In inclement conditions I think humans still out perform machines pretty handily. Eventually it'll be an even field or there will be sensors added to cars that can just see through the snow/rain but that's not where we are.

TwiztidK 11y ago

> Why would machines be worse than humans in bad weather if they are better than humans in good weather? [Because they are.](http://jalopnik.com/this-is-how-bad-self-driving-cars-suck-in-the-rain-1666268433 http://jalopnik.com/this-is-how-bad-self-driving-cars-suck-i...) Basically, the sensors they tend to equipped with aren't even close to being sufficient in poor whether. Humans are EXTREMELY good at pattern recognition and can manage to make it work, but computers just aren't at that level yet. It'll take a combination of improved sensors and computer vision (ie, pattern recognition) to make it happen.

greedo 11y ago

Because Google has mapped out the roads to a very specific level; this allows them to use sensors in the cars that don't have to have the fidelity of human eyes etc. As Google's maps become out of date, they need to be recreated or checked. Their system won't scale...