Consider the situation where data storage is virtually free.
Whenever I meet a friend I could copy ALL his movies and sound files. And it will be fast. And everybody will have 1000s or millions of media files. So its only an issue to get the really new stuff. And if anybody buys anything they can "give" it to all their friends. In such a system there is no longer any need for "the net".
When can we expect good computer vision? There are lots of progresses in the field. New statistical techniques. Faster algorithms for supervised learning. But still. I guess if you had asked 30 years ago when perceptrons were quite fashionable how long it would take to have real good computer vision you would have gotten the same estimate of 20 years. Doing some work in computer vision I must say that to my knowledge we are still very far from building anything thats real. We are rather at the stage where we discover 2 new problems for each problem solved. Problems are for example: Attention, efficient learning, efficient inference, symbol grounding, categorization. So I guess it will take many more years. Or forever.
What about self repair ? One of the really cool things about humans is that they mostly repair themselves. Our bodys endure constant abuse. Our bodies constantly repair the damage at least over approximately 100 years. A large number of robots would demand constant repairing.
Are robots really cheap? Lets face it people are there. We already have a very high rate of jobless people. Given the right taxation systems these people should be a lot less cheaper than any robot could ever be.
Dont get all of this wrong. Computer Vision and Robotics will improve. But it will improve the same way that tools improved throughout the history of mankind. They slowly get better and more useful. While we find novel ways of using them. And spend our time doing more interesting stuff. Like reading slashdot.
The appeal of genetic algorithms
on
Mutating Animations
·
· Score: 2, Informative
The field of machine learning consists of probably at least some 1000 researchers worldwide. Genetic Algorithms in the way they are usually used are an algorithm for optimization.
The problem of optimizing: You have a set of parameters p_i. You have a fitness/ energy or objective function F that somehow measures how good you are doing. The algorithm is supposed to yield some p_i that lead to optimal performance, that is maximal F.
What is a good algorithm? Whereas some years ago people would write sentences like "my algorithm is better than your algorithm" we now know that without prior knowledge about the function to be optimized all optimization algorithms are equally good. So today you can only say... my algorithm works better for this and that type of data because it incorporates this or that knowledge about the problem encountered.
Algorithms used Depending on what you know about the problem set you can then build various algorithms. The following are among the most common algorithms for optimization:
Methods based on gradients (i.e. vanilla, conjugate gradients),
Methods based on random search
and genetic Methods.
Why this is a good example for using genetic algorithms Genetic algorithms completely ignore the gradient information. In the case of walking the gradient of the fitness function may be difficult to define. That is why this is a classical example of using genetic algorithms.
Why genetic algorithms are absolutely hopeless for more interesting problems Genetic algorithms do not obtain a lot of information. In Particular for example David MacKay shows that the number of bits per generation is very low (1 bit for asexual, sqrt(size(genome)) for sexual systems). This means that the information acquired per generation is very low and that to obtain human like abilities it would take forever.
For interesting problems other optimization methods must be used that use information from the environment in a far more efficient way. Without learning I guess there can not be human like performance.
that already does quite cool stuff with short videos uploaded by virtually anyone.
If both are combined I could really imagine this to be useful. Imagine something like slashdot where editors select stories. Everybody would then sortof upload their clips that would get moderated. I dont see why this should not be possible.
Ok. Lets face it. Pattern recognition is improving slowly but steadily. We are now able to detect number plates at high speed. We can recognize people by their face or the way they walk. Not perfectly but every year algorithms improve a little bit.
In addition to that there are many promising algorithms out there that can for example learn what is surprising. So Pattern Recognition (parts of which where called AI some years ago) is getting there.
This will be exploited. And there is no way we can avoid that. As the technology evolves it starts to be possible to anyone to use it. Including the government. And they will use it to spy on us. Face it.
I think we will need to embrace this change. Forget privacy. That was the past. Given that the technolgy is there it will be used. The only thing we might be able to do is use the very same technology on those that use the technology on us.
So start gathering data on your MPs. Start to monitor how the data are used. Thats all we can do.
Two unknowns dont make stuff work
on
AI Going Nowhere?
·
· Score: 3, Insightful
I very much enjoy the works of Markram and Tsodyks. What they mainly analyze is how two nerve cells can interact with each other. They showed how they change their connection weights and how the timing of spikes, nerve impulses, affect how neurons connect to each other and how they transmit information.
While these studies tell us a lot about the underlying biology they do not tell us what these modes of information transmission are used for. For years it had been known that synapses have complex nonlinear properties. Biology pretty much does not constrain what functions neurons compute.
Thats why I do not believe that such studies will bring us nearer to real AI anytime soon. The algorithms coming from these systems are severely underconstrained. A lot of modelling has followed the pioneering works of Markram and Tsodyks, one of them being Maas. All these algorithms are very fascinating and might yield insight into the functioning of the nervous system.
The algorithms however are lightyears from being applicable to real world problems. The field of AI is old and in some sense quite mature. None of the "biologically inspired" algorithms today can compete with state of the art machine learning techniques.
The great thing about the recent development in so-called cognitive systems is that they start to address more real problems. The time of toy problems is over. It is not enough to just follow a line. Only the challenge from the real world can make algorithms in any way "clever" or meaningful.
This is why I find it truly inspiring that so much research is going into these systems these days.
Sadly however most of neuroscience these days is still far from these questions. Most electrophysiologists that for example study the visual system show it trivial stimuli such as bars or gratings. In some sense a system can only show its capability when the stimuli are rich enough.
Nevertheless there is clearly a move these days towards larger more interesting problems even in neuroscience. We should be inspired by the works of the roboticists.
I am not sure about this but there seems to be a certain marketing push behind the project. The description of whos supposed to download it is hilarious. But all the machine learning stuff is hidden behind buzzwords. Why do they not put up a description of the algorithm or at least about its rational. I am involved in machine learning myself and most of my colleagues are extremely careful when using words like "the brain". And there is a usually a strong anticorrelation between the quality of work and the use of such buzzwords.
We will clearly see more "intelligent" machines in the future. And the direction that current "artificial intelligence" is going this means that these machines will learn from what is out there.
This directly implies that the behavior of the machine will depend in a fuzzy way on the past "experience" of that machine. This however also means that we will not be able to predict exactly how it is behaving. Only in the way we can understand other peoples behavior that have also learned this behavior from the real world.
While these learning systems will make prediction difficult it will make explicit what the machine is trying to do through the learning process. While we wont know how a machine does "it" it will always present the right possible actions to us. Microsoft Word 21XX will clearly not need us to search menus if we want to change the formatting of the text.
Well... the the second law of thermodynamics seems to imply that even such micromachines are not possible because entropy would then be changed into the wrong direction.
As said on lectureonline.cl.msu.edu:(You can browse through the book changing the url. "These machine then violate the second law of thermodynamics, as we will see in the following, and are thus impossible to work. This is much harder to see, because the concepts are rather delicate. The book proceeds to introduce the concept of Entropy".
Other works discuss if the second law of thermodynamics (which forbids your machines) can be derived from quantum mechanics : Such quantum mechanical systems would be adequate to describe your micromachines.
I tried to design such micromachines too some day and I even think that scientific american had an article about it that I unfortunately cant find
I am a scientist now and after studying physics I guess I am completely cured from the idea that there could be a perpetuum mobile, a machine that produces energy out of vacuum.
But I remember say 20 years ago I spent a long time trying to invent such machines. I kept trying to design it and kept asking people why it wouldnt work. It took a long (frustrating) time before I could sortof acknowledge that it didnt seem to work.
So honestly... who has undergone the same process?
I agree that we might use different meanings of self-assembly. Yes for me PDP (Parallel Distributed Processing) is some kind of a self-assembly. And so are support vector machines and graphical Bayesian networks.
There are various conceptions of what Bayesian Models actually do. In some cases, e.g. Mixture models, you can easily interpret them in the sense that each model "tries to explain what it can" while at the same time interacting with other "agents" or models about which inputs it is responsible for. As opposed to the cited computer memory these elements actually interact. A typical example is the so-called "explaining away". Bayesian Models in many implementations furthermore do what is called "message passing". What is this if not interacting with other elements.
I however agree in the sense that the buzz-word "self-organizing" is overused. Often using these words grants publication of results that would otherwise not be worth a publication since it often makes boring results sound new.
p.s.: Since you seem to be working in the same field as I currently am. May I ask you who I am really discussing with?
It turns out that already today all successful applications of socalled "artificial intelligence" are self assembling.
In the first approaches to artificial intelligence people used programming languages to obtain systems that generate intelligent or at least apparently intelligent behavior.
All newer approaches to artificial intelligence start with a large number of very simple units that, learning from data from the real world, develop specific patterns of connections. Many models even develop their own structure in such a way.
From my perspective is intelligence as well as artificial intelligence only possible in a system that can self-structure.
While it is clearly true that only the recent advances in computer speed allowed the Computer Vision Systems we are seeing now there are also other important influences.
In particular there are really also better algorithms than a number of years ago. Many if not most successful computer vision systems use statistical Methods. In the case of faces for example they often build a probabilistic model of what a face is. Such models know that a face should usually has eyes but not always. That some people have beards etc. And these models train themselves up from a database of stimuli, for example real faces.
A number of recent advances makes such probabilistic models fast enough to work well on real world data. In a sense is the problem of computer vision very similar to the problem of understanding a voice or extracting the highest possible bitrate from a stream of data transmitted via a telephone line. And indeed the resulting algorithms are often surprisingly similar
There are plans for building a similar system to GPS in Europe so that we are not too much depending on the american empire.
The following
page nicely explains the concept. More is available here .
This is technically very interesting and should open up new possibilities for navigation. Furthermore being constructed jointly by many partners and nations we can be reasonably sure that it can not be compromised by one weak leader.
It would not be a good idea to release this medicine right at the moment. I completely agree with the firms decision.
Why you may ask. Antibiotics that are not supplied to the general public will be useful against ALL bacteria out there. All the ones that are acutally used start to develop species out there that no longer are sensitive to the antibiotics.
It might be important for the survival for many people to not use certain antibiotics to early.
As an example lets assume that some dangerous rogue state, say the united states develops this great biological weapon. Resistant against ALL known antibiotics. Then this drug might rescue many lives.
In switzerland we have been having a firm that has been promoting building such a train for a long time. Several universities have done feasibility studies to show that it can be done. It is currently discussed if the state should pay for the enormous costs involved.
The worrying thing is that geneticists seem to believe that genes define what we are. They search for genes for intelligence - and for preference for red whine.
My bet is that they will soon realize that speech underlying human culture is due to networks of thousands of genes and also due to advances in human culture and technology.
There can not possibly be a gene making a human out of a monkey.
You cant possibly hire enough people to fly each of these planes if you use them for surveillance. Working in an institute that is heavily involved in modern forms of AI I can assure you that the number of crashing/dying planes will be immense.
Its really difficult to make a driving robot come back home. They always hit things or are very slow.
These planes better be really cheap! And the firms that deliver them will have to deliver them in the millions if a few hundred of them are to be in the air at any point of time.
I guess my humanities friends that always told me that culture is about to turn us into a giant meta-organism are right after all.
Interesting that it took plenty of technology to get there and surprisingly little "humanities" moderation. But then they say that a meta-organism has been what we were all along.
Slashdot Mirror
Consider the situation where data storage is virtually free. Whenever I meet a friend I could copy ALL his movies and sound files. And it will be fast. And everybody will have 1000s or millions of media files. So its only an issue to get the really new stuff. And if anybody buys anything they can "give" it to all their friends. In such a system there is no longer any need for "the net".
Ok lets look at a number of problems
When can we expect good computer vision? There are lots of progresses in the field. New statistical techniques. Faster algorithms for supervised learning. But still. I guess if you had asked 30 years ago when perceptrons were quite fashionable how long it would take to have real good computer vision you would have gotten the same estimate of 20 years. Doing some work in computer vision I must say that to my knowledge we are still very far from building anything thats real. We are rather at the stage where we discover 2 new problems for each problem solved. Problems are for example: Attention, efficient learning, efficient inference, symbol grounding, categorization. So I guess it will take many more years. Or forever.
What about self repair ? One of the really cool things about humans is that they mostly repair themselves. Our bodys endure constant abuse. Our bodies constantly repair the damage at least over approximately 100 years. A large number of robots would demand constant repairing.
Are robots really cheap? Lets face it people are there. We already have a very high rate of jobless people. Given the right taxation systems these people should be a lot less cheaper than any robot could ever be.
Dont get all of this wrong. Computer Vision and Robotics will improve. But it will improve the same way that tools improved throughout the history of mankind. They slowly get better and more useful. While we find novel ways of using them. And spend our time doing more interesting stuff. Like reading slashdot.
The field of machine learning consists of probably at least some 1000 researchers worldwide. Genetic Algorithms in the way they are usually used are an algorithm for optimization.
The problem of optimizing: You have a set of parameters p_i. You have a fitness/ energy or objective function F that somehow measures how good you are doing. The algorithm is supposed to yield some p_i that lead to optimal performance, that is maximal F.
What is a good algorithm? Whereas some years ago people would write sentences like "my algorithm is better than your algorithm" we now know that without prior knowledge about the function to be optimized all optimization algorithms are equally good. So today you can only say ... my algorithm works better for this and that type of data because it incorporates this or that knowledge about the problem encountered.
Algorithms used Depending on what you know about the problem set you can then build various algorithms. The following are among the most common algorithms for optimization: Methods based on gradients (i.e. vanilla, conjugate gradients), Methods based on random search and genetic Methods.
Why this is a good example for using genetic algorithms Genetic algorithms completely ignore the gradient information. In the case of walking the gradient of the fitness function may be difficult to define. That is why this is a classical example of using genetic algorithms.
Why genetic algorithms are absolutely hopeless for more interesting problems Genetic algorithms do not obtain a lot of information. In Particular for example David MacKay shows that the number of bits per generation is very low (1 bit for asexual, sqrt(size(genome)) for sexual systems). This means that the information acquired per generation is very low and that to obtain human like abilities it would take forever.
For interesting problems other optimization methods must be used that use information from the environment in a far more efficient way. Without learning I guess there can not be human like performance.
Slashdot is a real big success story. The moderation system makes sure I see only at least remotely relevant or funny stuff.
Now checkout tv.oneworld.net
that already does quite cool stuff with short videos uploaded by virtually anyone.If both are combined I could really imagine this to be useful. Imagine something like slashdot where editors select stories. Everybody would then sortof upload their clips that would get moderated. I dont see why this should not be possible.
Ok. Lets face it. Pattern recognition is improving slowly but steadily. We are now able to detect number plates at high speed. We can recognize people by their face or the way they walk. Not perfectly but every year algorithms improve a little bit.
In addition to that there are many promising algorithms out there that can for example learn what is surprising. So Pattern Recognition (parts of which where called AI some years ago) is getting there.
This will be exploited. And there is no way we can avoid that. As the technology evolves it starts to be possible to anyone to use it. Including the government. And they will use it to spy on us. Face it.
I think we will need to embrace this change. Forget privacy. That was the past. Given that the technolgy is there it will be used. The only thing we might be able to do is use the very same technology on those that use the technology on us.
So start gathering data on your MPs. Start to monitor how the data are used. Thats all we can do.
I very much enjoy the works of Markram and Tsodyks. What they mainly analyze is how two nerve cells can interact with each other. They showed how they change their connection weights and how the timing of spikes, nerve impulses, affect how neurons connect to each other and how they transmit information.
While these studies tell us a lot about the underlying biology they do not tell us what these modes of information transmission are used for. For years it had been known that synapses have complex nonlinear properties. Biology pretty much does not constrain what functions neurons compute.
Thats why I do not believe that such studies will bring us nearer to real AI anytime soon. The algorithms coming from these systems are severely underconstrained. A lot of modelling has followed the pioneering works of Markram and Tsodyks, one of them being Maas. All these algorithms are very fascinating and might yield insight into the functioning of the nervous system.
The algorithms however are lightyears from being applicable to real world problems. The field of AI is old and in some sense quite mature. None of the "biologically inspired" algorithms today can compete with state of the art machine learning techniques.
The great thing about the recent development in so-called cognitive systems is that they start to address more real problems. The time of toy problems is over. It is not enough to just follow a line. Only the challenge from the real world can make algorithms in any way "clever" or meaningful.
This is why I find it truly inspiring that so much research is going into these systems these days.
Sadly however most of neuroscience these days is still far from these questions. Most electrophysiologists that for example study the visual system show it trivial stimuli such as bars or gratings. In some sense a system can only show its capability when the stimuli are rich enough.
Nevertheless there is clearly a move these days towards larger more interesting problems even in neuroscience. We should be inspired by the works of the roboticists.
I am not sure about this but there seems to be a certain marketing push behind the project. The description of whos supposed to download it is hilarious. But all the machine learning stuff is hidden behind buzzwords. Why do they not put up a description of the algorithm or at least about its rational. I am involved in machine learning myself and most of my colleagues are extremely careful when using words like "the brain". And there is a usually a strong anticorrelation between the quality of work and the use of such buzzwords.
We will clearly see more "intelligent" machines in the future. And the direction that current "artificial intelligence" is going this means that these machines will learn from what is out there.
This directly implies that the behavior of the machine will depend in a fuzzy way on the past "experience" of that machine. This however also means that we will not be able to predict exactly how it is behaving. Only in the way we can understand other peoples behavior that have also learned this behavior from the real world.
While these learning systems will make prediction difficult it will make explicit what the machine is trying to do through the learning process. While we wont know how a machine does "it" it will always present the right possible actions to us. Microsoft Word 21XX will clearly not need us to search menus if we want to change the formatting of the text.
Well ... the the second law of thermodynamics seems to imply that even such micromachines are not possible because entropy would then be changed into the wrong direction.
As said on lectureonline.cl.msu.edu:(You can browse through the book changing the url. "These machine then violate the second law of thermodynamics, as we will see in the following, and are thus impossible to work. This is much harder to see, because the concepts are rather delicate. The book proceeds to introduce the concept of Entropy".
Other works discuss if the second law of thermodynamics (which forbids your machines) can be derived from quantum mechanics : Such quantum mechanical systems would be adequate to describe your micromachines.
I tried to design such micromachines too some day and I even think that scientific american had an article about it that I unfortunately cant find
I am a scientist now and after studying physics I guess I am completely cured from the idea that there could be a perpetuum mobile, a machine that produces energy out of vacuum.
But I remember say 20 years ago I spent a long time trying to invent such machines. I kept trying to design it and kept asking people why it wouldnt work. It took a long (frustrating) time before I could sortof acknowledge that it didnt seem to work.
So honestly... who has undergone the same process?
I agree that we might use different meanings of self-assembly. Yes for me PDP (Parallel Distributed Processing) is some kind of a self-assembly. And so are support vector machines and graphical Bayesian networks.
There are various conceptions of what Bayesian Models actually do. In some cases, e.g. Mixture models, you can easily interpret them in the sense that each model "tries to explain what it can" while at the same time interacting with other "agents" or models about which inputs it is responsible for. As opposed to the cited computer memory these elements actually interact. A typical example is the so-called "explaining away". Bayesian Models in many implementations furthermore do what is called "message passing". What is this if not interacting with other elements.
I however agree in the sense that the buzz-word "self-organizing" is overused. Often using these words grants publication of results that would otherwise not be worth a publication since it often makes boring results sound new.
p.s.: Since you seem to be working in the same field as I currently am. May I ask you who I am really discussing with?
It turns out that already today all successful applications of socalled "artificial intelligence" are self assembling.
In the first approaches to artificial intelligence people used programming languages to obtain systems that generate intelligent or at least apparently intelligent behavior.
All newer approaches to artificial intelligence start with a large number of very simple units that, learning from data from the real world, develop specific patterns of connections. Many models even develop their own structure in such a way.
From my perspective is intelligence as well as artificial intelligence only possible in a system that can self-structure.
While it is clearly true that only the recent advances in computer speed allowed the Computer Vision Systems we are seeing now there are also other important influences.
In particular there are really also better algorithms than a number of years ago. Many if not most successful computer vision systems use statistical Methods. In the case of faces for example they often build a probabilistic model of what a face is. Such models know that a face should usually has eyes but not always. That some people have beards etc. And these models train themselves up from a database of stimuli, for example real faces.
A number of recent advances makes such probabilistic models fast enough to work well on real world data. In a sense is the problem of computer vision very similar to the problem of understanding a voice or extracting the highest possible bitrate from a stream of data transmitted via a telephone line. And indeed the resulting algorithms are often surprisingly similar
There are plans for building a similar system to GPS in Europe so that we are not too much depending on the american empire. The following page nicely explains the concept. More is available here . This is technically very interesting and should open up new possibilities for navigation. Furthermore being constructed jointly by many partners and nations we can be reasonably sure that it can not be compromised by one weak leader.
It would not be a good idea to release this medicine right at the moment. I completely agree with the firms decision.
Why you may ask. Antibiotics that are not supplied to the general public will be useful against ALL bacteria out there. All the ones that are acutally used start to develop species out there that no longer are sensitive to the antibiotics.
It might be important for the survival for many people to not use certain antibiotics to early.
As an example lets assume that some dangerous rogue state, say the united states develops this great biological weapon. Resistant against ALL known antibiotics. Then this drug might rescue many lives.
So I guess they shouldnt release it.
for obvious reasons
Mod this down please
No uncommented links to porn pictures on slashdot please. Its also offtopic
In switzerland we have been having a firm that has been promoting building such a train for a long time. Several universities have done feasibility studies to show that it can be done. It is currently discussed if the state should pay for the enormous costs involved.
After checking out this list of ads I was only wondering which firm would do it first.
I would have loved our IT classes if they proceeded like that
They tell you how easy it is to get a university degree for your lifetime's work and achievements. I keep wondering what I got my PhD for...
My bet is that they will soon realize that speech underlying human culture is due to networks of thousands of genes and also due to advances in human culture and technology.
There can not possibly be a gene making a human out of a monkey.You cant possibly hire enough people to fly each of these planes if you use them for surveillance. Working in an institute that is heavily involved in modern forms of AI I can assure you that the number of crashing /dying planes will be immense.
Its really difficult to make a driving robot come back home. They always hit things or are very slow.
These planes better be really cheap! And the firms that deliver them will have to deliver them in the millions if a few hundred of them are to be in the air at any point of time.
I guess my humanities friends that always told me that culture is about to turn us into a giant meta-organism are right after all. Interesting that it took plenty of technology to get there and surprisingly little "humanities" moderation.
But then they say that a meta-organism has been what we were all along.