4 ms·

> you shouldn’t have to harvest lots of data for intelligent AI. Er.... that's not how it works. As a Valve employee put it (a talk about ML recognising cheate

> you shouldn’t have to harvest lots of data for intelligent AI. Er.... that's not how it works. As a Valve employee put it (a talk about ML recognising cheaters), machine learning is just an equation: y = f(x) where you can actually derive the FUNCTION f if you have enough x/y pairs. Without those pairs then most of the successful techniques in machine learning fail to function. Yes you can generate your own data but generating realistic stuff is a huge can of worms.

paraschopra 8y ago
Not true. You can code your prior knowledge into the model, and make do with way less x/y pairs. Data is a substitute for doing engineering by hand. If you do the latter, you require less data.
- Quarrelsome 8y ago
  I don't see how it makes my statement untrue. The most effective ML techniques are more effective the more x/y pairs you have. Sure, you can get a leg up if you partially define F yourself but then you're no longer being as hands off as before. In practice the WHOLE of F() will likely be a blend of machine learning as well as heuristics as a product of various layers that are one or the other.
  - gugagore 8y ago
    Not the poster you're replying to, but it's just that you said "that's not how it works", and then suggested that the only way to get F is through data, even if it means generating synthetic data. So that's wrong and untrue, as you recognize in this latest comment by saying you can partially define F. To make a concrete example, the convolution operator, the number of layers, their connectivity--all of that partially defines F. Finding a structure is what allows you to learn from the x/y pairs instead of just storing them in a look-up table. It makes sense, then, that finding a better structure may allow you to learn from fewer x/y pairs. This is Apple's purported approach. It also means that if you have a huge number of x/y pairs, you might benefit from (or get away with) using a simpler structure. This is Google's approach in the domains where they have a lot of data. It is true that an important lesson has been "when designers build structure into machine learning systems, they often build in the wrong things, and they end up making commitments that the data cannot undo". But there's still a lot of structure that we put into machine learning system, and we're still learning about what good structures are for different applications, and different regimes of data.
- kthejoker2 8y ago
  Traffic is a chaos system, you can't engineer the edge cases. At best you have heuristics.
  - gugagore 8y ago
    I don't think "traffic is chaotic" is the issue with driverless cars. We'd be satisfied with a car that solved everything except for global (i.e. non-local) traffic flow issues. That said, a dynamical system may exhibit chaos for one setting of the parameters, and may not exhibit chaos for other settings of the parameters. It may be true that the equations governing traffic flow have the potential to exhibit chaos, and do exhibit chaos when 1. humans drive, or 2. the drivers have only local sensing. But that doesn't mean it would still be true for driverless cars. But again, that's not the issue at play. We wouldn't fault an driverless car for getting stuck in traffic. It's more like it has to work when it's snowing, when you can't see the lane markers, when a kid in a trashbag costume jumps into the street, etc. Those are the edge cases that we care about right now.
dullgiulio 8y ago
Plus, I don't quite get what is the privacy being violated when taking a lot of pictures of red traffic lights...
- simonh 8y ago
  Mine the exif data for the time and geolocation of the people taking the pictures and what camera/phone they used. You might even be able to tell which direction they were travelling in.
rjf72 8y ago
There are different ways of training things than relying other people's data. For instance alpha zero is still the most remarkable achievement in ML or 'AI' in my opinion, and it used 0 human games in learning to play chess. It training was entirely from self-play. I think models that rely on learning from these sort of weighted monte carlo type systems are likely to end up far more robust than ones that try to learn from humans. In the AlphaZero case this was provably true. But in the more general case, the optimal solution for humans and the optimal solution for machines is not necessarily the same. So in driving, for instance, it may be the case that operating a vehicle in one way is safest for a human - yet not necessarily so for a machine. It's not hard to come up with examples. We are awesome at intuition and awful at microsecond reactions. Computers are awful at intuition and awesome at microsecond reactions. The optimal style of 'play' is likely going to differ. Yet if the machine is trained based on human data it would end up trying to shoehorn itself into replicating human behavior which may be suboptimal for it. All that said, this sounds more like marketing talk. My guess is that Apple simply decided that self driving is extremely difficult and is going to come with some major growing pains which could negatively affected their crafted brand image, making it generally a poor investment of time and resources.
- Khanhanhan 8y ago
  Driving, unlike chess, is a game with no bounding rules and potentially tragic consequences.

paraschopra 8y ago

Not true. You can code your prior knowledge into the model, and make do with way less x/y pairs. Data is a substitute for doing engineering by hand. If you do the latter, you require less data.

Quarrelsome 8y ago
I don't see how it makes my statement untrue. The most effective ML techniques are more effective the more x/y pairs you have. Sure, you can get a leg up if you partially define F yourself but then you're no longer being as hands off as before. In practice the WHOLE of F() will likely be a blend of machine learning as well as heuristics as a product of various layers that are one or the other.
- gugagore 8y ago
  Not the poster you're replying to, but it's just that you said "that's not how it works", and then suggested that the only way to get F is through data, even if it means generating synthetic data. So that's wrong and untrue, as you recognize in this latest comment by saying you can partially define F. To make a concrete example, the convolution operator, the number of layers, their connectivity--all of that partially defines F. Finding a structure is what allows you to learn from the x/y pairs instead of just storing them in a look-up table. It makes sense, then, that finding a better structure may allow you to learn from fewer x/y pairs. This is Apple's purported approach. It also means that if you have a huge number of x/y pairs, you might benefit from (or get away with) using a simpler structure. This is Google's approach in the domains where they have a lot of data. It is true that an important lesson has been "when designers build structure into machine learning systems, they often build in the wrong things, and they end up making commitments that the data cannot undo". But there's still a lot of structure that we put into machine learning system, and we're still learning about what good structures are for different applications, and different regimes of data.
kthejoker2 8y ago
Traffic is a chaos system, you can't engineer the edge cases. At best you have heuristics.
- gugagore 8y ago
  I don't think "traffic is chaotic" is the issue with driverless cars. We'd be satisfied with a car that solved everything except for global (i.e. non-local) traffic flow issues. That said, a dynamical system may exhibit chaos for one setting of the parameters, and may not exhibit chaos for other settings of the parameters. It may be true that the equations governing traffic flow have the potential to exhibit chaos, and do exhibit chaos when 1. humans drive, or 2. the drivers have only local sensing. But that doesn't mean it would still be true for driverless cars. But again, that's not the issue at play. We wouldn't fault an driverless car for getting stuck in traffic. It's more like it has to work when it's snowing, when you can't see the lane markers, when a kid in a trashbag costume jumps into the street, etc. Those are the edge cases that we care about right now.

dullgiulio 8y ago

Plus, I don't quite get what is the privacy being violated when taking a lot of pictures of red traffic lights...

simonh 8y ago
Mine the exif data for the time and geolocation of the people taking the pictures and what camera/phone they used. You might even be able to tell which direction they were travelling in.

rjf72 8y ago

There are different ways of training things than relying other people's data. For instance alpha zero is still the most remarkable achievement in ML or 'AI' in my opinion, and it used 0 human games in learning to play chess. It training was entirely from self-play. I think models that rely on learning from these sort of weighted monte carlo type systems are likely to end up far more robust than ones that try to learn from humans. In the AlphaZero case this was provably true. But in the more general case, the optimal solution for humans and the optimal solution for machines is not necessarily the same. So in driving, for instance, it may be the case that operating a vehicle in one way is safest for a human - yet not necessarily so for a machine. It's not hard to come up with examples. We are awesome at intuition and awful at microsecond reactions. Computers are awful at intuition and awesome at microsecond reactions. The optimal style of 'play' is likely going to differ. Yet if the machine is trained based on human data it would end up trying to shoehorn itself into replicating human behavior which may be suboptimal for it. All that said, this sounds more like marketing talk. My guess is that Apple simply decided that self driving is extremely difficult and is going to come with some major growing pains which could negatively affected their crafted brand image, making it generally a poor investment of time and resources.

Khanhanhan 8y ago
Driving, unlike chess, is a game with no bounding rules and potentially tragic consequences.