When human beings see something unexpected, we do a double take
Of course, you first need to see something unexpected. In the famous video of white/black people passing a ball, very few people noticed the gorilla. They never did a double take. https://www.youtube.com/watch?... This happens all the time in real life.
For example: there are two stable isotopes of lithium. Chemically they are identical,
No, they are not. For example, one of the methods of separating them is the COLEX process https://en.wikipedia.org/wiki/... which exploits their different chemical properties.
Searle's Chinese Room is one of the stupidest ideas ever proposed.
That said, you're not even applying it correctly. The premise of the Chinese Room is that the room produces behavior indistinguishable from a real Chinese speaker. Any time you can point to a failure of an AI system, it is clearly violating that premise.
Google, having more hardware than any research department. Google, having more man-hours to spend than any research department. Google, having more data than all research departments put together.
Even if we assume all of that is true, it still makes sense for them to release their translation server somewhere between the moment it was better than the old one, and before it was 100% perfect.
The only reason for the previous AI winter was the fact that the AI at that time could not be monetized. We are way beyond that now. AI is making profit, and therefore there is continued effort to improve it and make even more money.
The problem with any machine vision recognition system is that there is really no way of looking at a still image (or even a still scene with depth) and/knowing/ which Objects are discrete and independent
This is not a problem. If you have enough data, the machine will find the patterns.
And if it has a window, how is that different from a picture on the wall?
A window usually lets in light, whereas a picture doesn't. Also, the perspective is usually different, as well as the type of objects you can see in a picture vs through a window.
Yes, they actually learn to generalize. This can be tested fairly simply by showing them unique images that differ significantly from any of the training images, and seeing how many get classified correctly.
The question (which the writer didn't ask or answer) is how the machine learning systems can be improved to be more resistant against such simple modifications.
Because fluctuations in external factors could make it lucrative to invest in a bunch of miners, which would allow a 51% attack on the bitcoin network. If the miners are not useful for anything except bitcoin, then such an attack would leave all your miners worthless the instant you performed it. In most scenarios that would not be attractive. If you did all the investments for such a large bitcoin miner, it would benefit you to just mine bitcoins for a prolonged period of time without attacking the network.
Most of the stuff we use copper for (like wiring in houses), gold actually works better,
No, copper conducts electricity better than gold. Only silver is slightly better than copper. In addition, copper only weighs half of gold, has slightly higher melting point, and is twice as strong, making it overall much superior for wiring than gold. The big reason for using gold instead of copper in some electrical applications is that gold doesn't tarnish, but for most cases, you only need a thin gold plating to benefit from that.
Either way they're crunching numbers. Might as well crunch something for all our benefit.
Strangely enough, you don't want bitcoin mining to be useful for any other purpose, otherwise profit motives for the other purpose could make bitcoin network less secure.
Either way, if folding@home is beneficial, then it's worth something, and people should be getting paid for the effort.
You're right - it doesn't grow with time, it grows with transactions. It has nothing to do with the mining capacity,
The computational complexity is adjusted based on mining speed so that there is (on average) 1 block mined every 10 minutes. If mining capacity goes up, then more blocks are found, and difficulty is increased. If mining capacity goes down, difficulty is decreased.
The amount of transactions only affects complexity indirectly. As there is more demand for transactions, the mining fee goes up (not linearly, because the fixed block reward is much more dominant right now), which makes mining more attractive, and as mining gets more attractive, the capacity increases, which increases difficulty.
It has numerous applications which have nothing to do with how shiny it is.
But only 10% of the mined gold is actually used for those applications. The rest is stored in vaults or converted to jewelry, for no other reason than that's it's shiny and perceived to be valuable.
How many millions of people on this planet are on the edge of starvation? And you seriously propose increasing the cost of energy?
It's not a problem if you lower other taxes by the same amount.
You're proposing something that could literally kill millions? If not more? Because "the cost for energy should include its true cost, which includes the cost of repairing the damage it cases"?!?!?
You overlook the fact that people on the edge of starvation will also suffer when the effects of the damage come back to them.
Since the computational complexity of the BC transactions grows with time
That's not how it works. The complexity grows with the mining capacity. And mining capacity grows (and possibly shrinks) with price of bitcoin, mining reward, and electricity. The mining reward consists of a fixed reward per block (halved every 4 years), plus a fee per transaction (determined by market mechanism)
I have to wonder, are these AIs really learning to generalize? Or are they having millions of examples, so many that any live image must be within a tiny delta of a known image?
This is a well known problem, called "overfitting". You can test it by only training your model on a subset of the images, and then see how it does on the remaining ones. There are various techniques to combat overfitting.
Manure burns because cellulose is very hard to digest.
high fat high protein diet == low obsesity
That's mostly because people eat less on such a diet. It makes you feel full for longer. Try eating a block of cheese, and then watch your poop the next day. Most likely it is perfectly normal, indicating that pretty much all of the fat was absorbed.
high carb diet == obsesity
Also not true. Plenty of people around the world eat (or ate) a high carb diet and are perfectly lean.
The things that make you fat are usually foods that are both sweet and fat. Try eating a bowl of plain sugar. It's disgusting. Try eating a bowl of plain cream. Not very appetizing either. Now mix them together, chill them, and you have ice cream. All of a sudden, you can eat both bowls.
No, that analogy is flawed. When reverse engineering you ALWAYS start with something that ALREADY works.
The analogy was about the requirement to have brains built out of organic matter. You're talking about something else. Next time, try to address the actual argument.
Trying to use a Linear process to understand a Non-Linear system will never work.
That's why all neural models are non-linear.
Without a way to MEASURE it, HOW do you know if what you are doing is moving towards or away from the goal post???
Quite simple. You just focus on the behavior. You can measure the inputs and outputs, and if they get closer to the behavior of a real brain, you know you're getting closer to the goal post. That's how evolution shaped our brain after all, simply by looking at the outputs and see if they benefit survival and reproduction.
Slashdot Mirror
When human beings see something unexpected, we do a double take
Of course, you first need to see something unexpected. In the famous video of white/black people passing a ball, very few people noticed the gorilla. They never did a double take. https://www.youtube.com/watch?... This happens all the time in real life.
For example: there are two stable isotopes of lithium. Chemically they are identical,
No, they are not. For example, one of the methods of separating them is the COLEX process https://en.wikipedia.org/wiki/... which exploits their different chemical properties.
Searle's Chinese Room is one of the stupidest ideas ever proposed.
That said, you're not even applying it correctly. The premise of the Chinese Room is that the room produces behavior indistinguishable from a real Chinese speaker. Any time you can point to a failure of an AI system, it is clearly violating that premise.
Google, having more hardware than any research department.
Google, having more man-hours to spend than any research department.
Google, having more data than all research departments put together.
Even if we assume all of that is true, it still makes sense for them to release their translation server somewhere between the moment it was better than the old one, and before it was 100% perfect.
So, translating a sentence from one language to another does not involve knowledge or skill ?
The only reason for the previous AI winter was the fact that the AI at that time could not be monetized. We are way beyond that now. AI is making profit, and therefore there is continued effort to improve it and make even more money.
I think the statement is more that ML systems use the wrong approach to identifying reality and get a very fragile performance as a result.
Yes, that's the writer's hunch, but nowhere does she show why we need a different approach rather than an improved version of the current one.
I doubt Google translate is that close to state-of-the-art.
Thing about it. Is there any reason for it to not be, any reason at all?
Hardware, time and data.
The problem with any machine vision recognition system is that there is really no way of looking at a still image (or even a still scene with depth) and /knowing/ which Objects are discrete and independent
This is not a problem. If you have enough data, the machine will find the patterns.
And if it has a window, how is that different from a picture on the wall?
A window usually lets in light, whereas a picture doesn't. Also, the perspective is usually different, as well as the type of objects you can see in a picture vs through a window.
Yes, they actually learn to generalize. This can be tested fairly simply by showing them unique images that differ significantly from any of the training images, and seeing how many get classified correctly.
The question (which the writer didn't ask or answer) is how the machine learning systems can be improved to be more resistant against such simple modifications.
Why would that make it less secure?
Because fluctuations in external factors could make it lucrative to invest in a bunch of miners, which would allow a 51% attack on the bitcoin network. If the miners are not useful for anything except bitcoin, then such an attack would leave all your miners worthless the instant you performed it. In most scenarios that would not be attractive. If you did all the investments for such a large bitcoin miner, it would benefit you to just mine bitcoins for a prolonged period of time without attacking the network.
Most of the stuff we use copper for (like wiring in houses), gold actually works better,
No, copper conducts electricity better than gold. Only silver is slightly better than copper. In addition, copper only weighs half of gold, has slightly higher melting point, and is twice as strong, making it overall much superior for wiring than gold. The big reason for using gold instead of copper in some electrical applications is that gold doesn't tarnish, but for most cases, you only need a thin gold plating to benefit from that.
That's true, but that doesn't mean that gold has zero intrinsic value, only that the intrinsic value is much lower than the current market price.
Correct. And the market price has been much higher than its 'intrinsic' value for as long as gold has existed.
Either way they're crunching numbers. Might as well crunch something for all our benefit.
Strangely enough, you don't want bitcoin mining to be useful for any other purpose, otherwise profit motives for the other purpose could make bitcoin network less secure.
Either way, if folding@home is beneficial, then it's worth something, and people should be getting paid for the effort.
You're right - it doesn't grow with time, it grows with transactions. It has nothing to do with the mining capacity,
The computational complexity is adjusted based on mining speed so that there is (on average) 1 block mined every 10 minutes. If mining capacity goes up, then more blocks are found, and difficulty is increased. If mining capacity goes down, difficulty is decreased.
The amount of transactions only affects complexity indirectly. As there is more demand for transactions, the mining fee goes up (not linearly, because the fixed block reward is much more dominant right now), which makes mining more attractive, and as mining gets more attractive, the capacity increases, which increases difficulty.
It has numerous applications which have nothing to do with how shiny it is.
But only 10% of the mined gold is actually used for those applications. The rest is stored in vaults or converted to jewelry, for no other reason than that's it's shiny and perceived to be valuable.
Thanks for your insights, but that was not an answer to my question.
How many millions of people on this planet are on the edge of starvation? And you seriously propose increasing the cost of energy?
It's not a problem if you lower other taxes by the same amount.
You're proposing something that could literally kill millions? If not more? Because "the cost for energy should include its true cost, which includes the cost of repairing the damage it cases"?!?!?
You overlook the fact that people on the edge of starvation will also suffer when the effects of the damage come back to them.
Please tell me you understand the difference? You can make an airplane with aluminium, you can't make shit with Bitcoin.
Then explain why we are continuing to mine gold, when there are already huge reserves above ground sitting in vaults and in people's jewelry.
Since the computational complexity of the BC transactions grows with time
That's not how it works. The complexity grows with the mining capacity. And mining capacity grows (and possibly shrinks) with price of bitcoin, mining reward, and electricity. The mining reward consists of a fixed reward per block (halved every 4 years), plus a fee per transaction (determined by market mechanism)
I have to wonder, are these AIs really learning to generalize? Or are they having millions of examples, so many that any live image must be within a tiny delta of a known image?
This is a well known problem, called "overfitting". You can test it by only training your model on a subset of the images, and then see how it does on the remaining ones. There are various techniques to combat overfitting.
Manure burns because cellulose is very hard to digest.
high fat high protein diet == low obsesity
That's mostly because people eat less on such a diet. It makes you feel full for longer. Try eating a block of cheese, and then watch your poop the next day. Most likely it is perfectly normal, indicating that pretty much all of the fat was absorbed.
high carb diet == obsesity
Also not true. Plenty of people around the world eat (or ate) a high carb diet and are perfectly lean.
The things that make you fat are usually foods that are both sweet and fat. Try eating a bowl of plain sugar. It's disgusting. Try eating a bowl of plain cream. Not very appetizing either. Now mix them together, chill them, and you have ice cream. All of a sudden, you can eat both bowls.
It's a lovely graph but it ends in 2012.
Here's one that has a few extra years:
https://en.wikipedia.org/wiki/...
No, that analogy is flawed. When reverse engineering you ALWAYS start with something that ALREADY works.
The analogy was about the requirement to have brains built out of organic matter. You're talking about something else. Next time, try to address the actual argument.
Trying to use a Linear process to understand a Non-Linear system will never work.
That's why all neural models are non-linear.
Without a way to MEASURE it, HOW do you know if what you are doing is moving towards or away from the goal post???
Quite simple. You just focus on the behavior. You can measure the inputs and outputs, and if they get closer to the behavior of a real brain, you know you're getting closer to the goal post. That's how evolution shaped our brain after all, simply by looking at the outputs and see if they benefit survival and reproduction.