From Merriam-Webster
cog·ni·tion
noun \käg-ˈni-shən\
From the MIT Cognitive Machines lab:
Cognitive Machines aims to:
(1) Create autonomous systems including interactive physical robots and synthetic characters in virtual worlds that learn to communicate in human-like ways;
(2) Understand how children learn to communicate through longitudinal in vivo observation and analysis;
(3) Develop tools for visualizing, searching, and analyzing large corpora (e.g., video) using task-dependent semantic models.
The idea of thinking machines has been around for thousands of years. It can be argued that the invention of computers was driven, in large part, by the idea of thinking machines.
Almost as soon as we figured out the basics of making computational machines, we turned to the business of making cognitive machines. The science of cognitive machines is sometimes called artificial intelligence. The field of artificial intelligence was founded in 1956 by a group of cognitive scientists at Dartmouth University. They began with the newly developing field of computer science. Much of their early work was based on Alan Turing's theories of computation and the Turing machine.
As the mechanics of creating fast, powerful, inexpensive, and widely available computers progressed, discoveries in the fields of human cognition and neuroscience began to reveal a vast gap between the workings of a modern computer and those of the human brain. Turing's theory of computation says, among other things, that a binary system, ones and zeroes, can solve any conceivable mathematical operation. All computers are built around that theory. Computer memory of all types is built to store ones and zeroes. However, as neuroscience has discovered, human memory is not binary.
One huge difference between computers and brains is determinism, and here is where cognitive theory and computer science start to tread on ground held by philosophy. A computer is completely deterministic. That is, for the conditions of its operation, i.e. its program and its inputs, there is only one possible outcome. Normally functioning computers don't make mistakes. Ever. Normally functioning people who write computer programs make mistakes all the time. Human cognition is completely non-deterministic. When you and I look at a photograph, we can come to completely different, perfectly valid conclusions about it. When you and I are given a set of requirements, we can produce different, perfectly valid programs to fulfill them. There is a class of computer program that can generate other computer programs. Ten computers running the same program will always produce the same code. A computer program is basically a list of things for the computer's processor to do. At its very most basic, the computer program tells the ones when to be ones, and the zeroes when to be zeroes. It will always do that the same way, every time.
Understanding and learning are major facets of cognition. A computer's understanding of its environment is completely deterministic. For all of its inputs at a given time, there is only one possible outcome. However, a computer is capable of simulating a non-deterministic understanding. If the set of inputs is sufficiently large, a deterministic program can appear to come to a non-deterministic outcome. When a computer program is written in a modern, high level language, the programmer uses letters and numbers and symbols arranged in such a way that it is (relatively) easy for other programmers to understand. However, anything other than lists of ones and zeroes are not immediately understandable by the computer. The computer requires either an interpreter or a compiler to transform the human readable program into something it understands. Because the number of possibilities in the grammar of the language, the computer uses a deterministic program that appears to come to a non-deterministic conclusion.
Since the beginning, designers and builders have said (tongue firmly in cheek) that a computer learns. They say that it has to learn to run the program it was given. The trouble is, it has to learn the same program every time it runs it. In human behavior, that is colloquially known as being an idiot. While human learning is far from perfect, computer learning is nonexistent. However, computers can simulate learning. Here is a straightforward example. One of the scourges of the computing era is junk email, or spam. All email programs now have spam filters that throw the unwanted email into a junk folder. These spam filters are adaptive. That is, they "learn" what junk mail looks like so they can separate the wheat from the chaff. What actually happens is that they start with a basic set of rules, like looking for the word, Viagra in the subject line. Over time, they add to those rules as they experience the various forms of incoming spam, like spelling Viagra with a "1" instead of an "i". So, you might ask, what is the difference between that behavior and actual learning. There is more to learning than just acquiring new information. Learning allows us to combine information we already have in myriad different ways to create new information. Computer learning is a close approximation to human learning, but in the end, it is not the same because it is still deterministic.
In the 1940s a concept known as neural networks was developed. The neural network is an attempt to simulate the working of the nervous systems of animals, including humans. Neural networks leave the realm of binary computation behind. The network consists of a set of input nodes, or neurons, that are applied to a set of internal neurons that contain various heuristic (rule based) weighting algorithms to apply the result to a third set of neurons, the output. Over time, the weighting algorithm is modified based on not only the input, but the output. In its time, the neural network was the closest thing to actual machine learning.
Artificial intelligence began as a theoretical exercise, but like everything else in computing, commercial demands have brought it out of the lab. The fields of robotics and pattern recognition have advanced artificial intelligence farther in the last ten years than in the previous forty. If you use Facebook, you have experienced a form of artificial intelligence. When you upload a picture containing a group of people, Facebook will find their faces and highlight them, asking if you want to tag them. That is actually a pretty sophisticated piece of software. The speed of your computer makes it appear to happen almost instantly. The massive computing power available these days pushes computers to the point that they are almost as fast as the human brain.
That's right. Almost. The human brain is capable of processing millions of simultaneously incoming events nearly instantaneously. The human body is an immensely complex system. The simple act of seeing the baseball heading for you and bringing millions of muscle and connective tissue cells to bear to bring the glove up to catch it is something the most advanced robotic systems can attempt only under controlled conditions. There will come a day, though, that a computer can simulate human thought, simply because of sheer computing power, It will be ugly and unsophisticated, not even approaching the perfection of human thought, but it might be close enough.
Next up: The computational universe.
: conscious mental activities : the activities of thinking, understanding, learning, and remembering
From the MIT Cognitive Machines lab:
Cognitive Machines aims to:
(1) Create autonomous systems including interactive physical robots and synthetic characters in virtual worlds that learn to communicate in human-like ways;
(2) Understand how children learn to communicate through longitudinal in vivo observation and analysis;
(3) Develop tools for visualizing, searching, and analyzing large corpora (e.g., video) using task-dependent semantic models.
The idea of thinking machines has been around for thousands of years. It can be argued that the invention of computers was driven, in large part, by the idea of thinking machines.
Almost as soon as we figured out the basics of making computational machines, we turned to the business of making cognitive machines. The science of cognitive machines is sometimes called artificial intelligence. The field of artificial intelligence was founded in 1956 by a group of cognitive scientists at Dartmouth University. They began with the newly developing field of computer science. Much of their early work was based on Alan Turing's theories of computation and the Turing machine.
As the mechanics of creating fast, powerful, inexpensive, and widely available computers progressed, discoveries in the fields of human cognition and neuroscience began to reveal a vast gap between the workings of a modern computer and those of the human brain. Turing's theory of computation says, among other things, that a binary system, ones and zeroes, can solve any conceivable mathematical operation. All computers are built around that theory. Computer memory of all types is built to store ones and zeroes. However, as neuroscience has discovered, human memory is not binary.
One huge difference between computers and brains is determinism, and here is where cognitive theory and computer science start to tread on ground held by philosophy. A computer is completely deterministic. That is, for the conditions of its operation, i.e. its program and its inputs, there is only one possible outcome. Normally functioning computers don't make mistakes. Ever. Normally functioning people who write computer programs make mistakes all the time. Human cognition is completely non-deterministic. When you and I look at a photograph, we can come to completely different, perfectly valid conclusions about it. When you and I are given a set of requirements, we can produce different, perfectly valid programs to fulfill them. There is a class of computer program that can generate other computer programs. Ten computers running the same program will always produce the same code. A computer program is basically a list of things for the computer's processor to do. At its very most basic, the computer program tells the ones when to be ones, and the zeroes when to be zeroes. It will always do that the same way, every time.
Understanding and learning are major facets of cognition. A computer's understanding of its environment is completely deterministic. For all of its inputs at a given time, there is only one possible outcome. However, a computer is capable of simulating a non-deterministic understanding. If the set of inputs is sufficiently large, a deterministic program can appear to come to a non-deterministic outcome. When a computer program is written in a modern, high level language, the programmer uses letters and numbers and symbols arranged in such a way that it is (relatively) easy for other programmers to understand. However, anything other than lists of ones and zeroes are not immediately understandable by the computer. The computer requires either an interpreter or a compiler to transform the human readable program into something it understands. Because the number of possibilities in the grammar of the language, the computer uses a deterministic program that appears to come to a non-deterministic conclusion.
Since the beginning, designers and builders have said (tongue firmly in cheek) that a computer learns. They say that it has to learn to run the program it was given. The trouble is, it has to learn the same program every time it runs it. In human behavior, that is colloquially known as being an idiot. While human learning is far from perfect, computer learning is nonexistent. However, computers can simulate learning. Here is a straightforward example. One of the scourges of the computing era is junk email, or spam. All email programs now have spam filters that throw the unwanted email into a junk folder. These spam filters are adaptive. That is, they "learn" what junk mail looks like so they can separate the wheat from the chaff. What actually happens is that they start with a basic set of rules, like looking for the word, Viagra in the subject line. Over time, they add to those rules as they experience the various forms of incoming spam, like spelling Viagra with a "1" instead of an "i". So, you might ask, what is the difference between that behavior and actual learning. There is more to learning than just acquiring new information. Learning allows us to combine information we already have in myriad different ways to create new information. Computer learning is a close approximation to human learning, but in the end, it is not the same because it is still deterministic.
In the 1940s a concept known as neural networks was developed. The neural network is an attempt to simulate the working of the nervous systems of animals, including humans. Neural networks leave the realm of binary computation behind. The network consists of a set of input nodes, or neurons, that are applied to a set of internal neurons that contain various heuristic (rule based) weighting algorithms to apply the result to a third set of neurons, the output. Over time, the weighting algorithm is modified based on not only the input, but the output. In its time, the neural network was the closest thing to actual machine learning.
Artificial intelligence began as a theoretical exercise, but like everything else in computing, commercial demands have brought it out of the lab. The fields of robotics and pattern recognition have advanced artificial intelligence farther in the last ten years than in the previous forty. If you use Facebook, you have experienced a form of artificial intelligence. When you upload a picture containing a group of people, Facebook will find their faces and highlight them, asking if you want to tag them. That is actually a pretty sophisticated piece of software. The speed of your computer makes it appear to happen almost instantly. The massive computing power available these days pushes computers to the point that they are almost as fast as the human brain.
That's right. Almost. The human brain is capable of processing millions of simultaneously incoming events nearly instantaneously. The human body is an immensely complex system. The simple act of seeing the baseball heading for you and bringing millions of muscle and connective tissue cells to bear to bring the glove up to catch it is something the most advanced robotic systems can attempt only under controlled conditions. There will come a day, though, that a computer can simulate human thought, simply because of sheer computing power, It will be ugly and unsophisticated, not even approaching the perfection of human thought, but it might be close enough.
Next up: The computational universe.
No comments:
Post a Comment