Science, mathematics, intelligence, every knowledge can be reduced to the ability of predicting the outcome in two directions: "what happens if certain conditions occurs?", "how can I make to obtain a certain outcome". Basically, this is all the expected benefit of knowledge. The theories are just models that have this unique purpose: to link certain pre-conditions to certain outcome.
Intelligence, in particular, is used to predict what happens in certain conditions (forward, like f(x)=?), along with finding the way to obtain a certain expected result (backward prediction), like "what x will make f(x)=y".
This might seem to be a too simplistic view of intelligence, as intelligence also contains a creative part - that solves very tricky puzzles that even computers cannot solve. Still, I would argue that this activity is also a backward-prediction applied in many steps. Let's consider o detective that is trying to solve a murder case. His problem is also a backward-prediction problem. He has a crime (the outcome), he tries to calculate the initial conditions that resulted in the outcome (crime). He needs to predict: "what is the most likely reason for this crime"; "who is most likely to have the reasons and opportunity?"; "what would have happen is this testimony would be true?"; "what other tests (questions) are likely to clarify the path that lead to this murder". The final outcome is clearly a backward-prediction problem, while the methods to obtain it are also, in my view, based on predictions (forward and backward).
Mathematics
Prediction is not the best word that would be used to describe mathematics (except in statistics maybe). Still, we can observe that theorems are really predictions of a certain outcome from certain pre-conditions (hypothesis).
The Pythagorean theorem predicts that the square of the hypotenuse is the sum of the squares of the other two sides of a right triangle. This is not the likely outcome, the mathematics can show very exact predictions that happens always - given some axioms. Predicting the same thing for a real object might not work exactly the same, as the measurements might be a bit off, the sides of the triangle are not exactly straight or the angle is not perfectly 90 degrees. Still, given the approximation we have, the mathematical predictions are accurate enough to be of great use to predict the length of the hypotenuse of something resembling a right triangle, only knowing the other 2 sides.
Theorems about number are also predictions, very exact ones as long as we talk about numbers. A type of prediction is found even in the simplest arithmetic - think associativity. Going more complex, you can predict that a polynomial equation of degree N have maximum N roots. Sometimes two roots are equal, but we still count them as being distinct. Sometimes the roots are "complex numbers, like i in "i^2 = -1", and we accept such roots only to be able to extend known methods from the real numbers when the intermediary results cannot be real numbers.
Not everything is the same in reality, for example you cannot divide matter at infinity as you can do with any small interval of real numbers. Even the length of a circle (2 * PI * Radius) is not something as exact as the PI number is; there are always two sides of a circle made from matter, and the inner length is slightly smaller that the outer one. There is no real perfect circle anyway. The earth is not a circle (more like an onion), the Sun has his irregularities. Still, the mathematical prediction is the best prediction we can make and good enough for practical problems involving round enough objects.
Physics
Think atomic reactions, even the nuclear bomb. You need a very well defined model to predict when we reach the critical mass to start a chain reaction. Still, the model is not perfect, as the involved matter is not as pure as the model requires, and also the measurements (energy, atom dimensions) are really measurements approximations. After one year of storage, the mass of the nuclear bomb is not the same anyway. I'm sure there are really big error intervals that are taken into account when applying any "exact" physics prediction. Still, I cannot think about any physics model that don't have as goal, ultimately, to do predictions about reality. Even string theory is a model that ultimately aims to make better and more exact predictions about reality than the relativity theory and quantum theory.
History
History is mainly a backward prediction. We have some observed realities (archaeological objects, stories, documents) and we try to backward predict the string of event that likely created the observed state. We cannot be sure that the documents are not lying, we just do our best to predict the most likely events that would create the history artifacts we can observe now. We cannot go back in time, we just backward-predict. Sometimes we forward predict based on hypothesis about past events, just to eliminate some events that proves to have consequences that do not harmonize with other likely backward predictions about the past. Predictions...
Human knowledge
The human life is a struggle to predict the likely outcome of certain current events, and find the best actions that have the desired outcome. We predict with good reliability who's face we see (while we can sometimes confuse people), we predict where our keys are (while sometimes they are somewhere else), we predict what is the best time to wake up to arrive at work at a reasonable hour.
Everything is forward and backward prediction, trying to optimize some internal objective functions, that are also weak approximations of what is good for us. For example we can do a very good prediction about the optimal way to obtain that smoking cigars that can damage our health. We predict, therefore we exist.
Artificial intelligence
For now artificial intelligence (or "machine learning") is mainly about predicting similarly with what human perception does. Andrew Ng told us a "rule of thumb" that the current "machine learning" is likely to be able to do what the human brain is able to do in less than 1 second, mainly automatic perception.
The computer can also solve a number of real problems that are hard for the human brain, but only when these problems can be formalized in a very specialized model, like in graph theory. The computer can play chess and recently beat even the best human players. Still, the computer can have a hard time to understand some language subtleties. The computer might still be unable to do some cognitive tasks in situations when even a child can do a good job.
The future of machine learning as I see it is better prediction about real life situations. These situations will not be exactly predicted, as the real life cannot be known with accuracy. Even a medical doctor diagnostic is often proven wrong, because we cannot have full knowledge about a patient. In order be become intelligent, the machine should be able to correlate a lot of statistical predictions about the regularities of the world in order to be able to do forward and backward prediction about real life.
This includes predictions about the language regularities, that should be correlated with likely scenario in life. Ambiguities like "the object does not fit in the bag because it is too small" can only be solved by also using some predictions about the behavior of these objects in real life.
Dear reader, please leave a message if you exist! ;) Also, please share this article if you find it interesting. Thank you.
Intelligence, in particular, is used to predict what happens in certain conditions (forward, like f(x)=?), along with finding the way to obtain a certain expected result (backward prediction), like "what x will make f(x)=y".
This might seem to be a too simplistic view of intelligence, as intelligence also contains a creative part - that solves very tricky puzzles that even computers cannot solve. Still, I would argue that this activity is also a backward-prediction applied in many steps. Let's consider o detective that is trying to solve a murder case. His problem is also a backward-prediction problem. He has a crime (the outcome), he tries to calculate the initial conditions that resulted in the outcome (crime). He needs to predict: "what is the most likely reason for this crime"; "who is most likely to have the reasons and opportunity?"; "what would have happen is this testimony would be true?"; "what other tests (questions) are likely to clarify the path that lead to this murder". The final outcome is clearly a backward-prediction problem, while the methods to obtain it are also, in my view, based on predictions (forward and backward).
Mathematics
Prediction is not the best word that would be used to describe mathematics (except in statistics maybe). Still, we can observe that theorems are really predictions of a certain outcome from certain pre-conditions (hypothesis).
The Pythagorean theorem predicts that the square of the hypotenuse is the sum of the squares of the other two sides of a right triangle. This is not the likely outcome, the mathematics can show very exact predictions that happens always - given some axioms. Predicting the same thing for a real object might not work exactly the same, as the measurements might be a bit off, the sides of the triangle are not exactly straight or the angle is not perfectly 90 degrees. Still, given the approximation we have, the mathematical predictions are accurate enough to be of great use to predict the length of the hypotenuse of something resembling a right triangle, only knowing the other 2 sides.
Theorems about number are also predictions, very exact ones as long as we talk about numbers. A type of prediction is found even in the simplest arithmetic - think associativity. Going more complex, you can predict that a polynomial equation of degree N have maximum N roots. Sometimes two roots are equal, but we still count them as being distinct. Sometimes the roots are "complex numbers, like i in "i^2 = -1", and we accept such roots only to be able to extend known methods from the real numbers when the intermediary results cannot be real numbers.
Not everything is the same in reality, for example you cannot divide matter at infinity as you can do with any small interval of real numbers. Even the length of a circle (2 * PI * Radius) is not something as exact as the PI number is; there are always two sides of a circle made from matter, and the inner length is slightly smaller that the outer one. There is no real perfect circle anyway. The earth is not a circle (more like an onion), the Sun has his irregularities. Still, the mathematical prediction is the best prediction we can make and good enough for practical problems involving round enough objects.
Physics
Think atomic reactions, even the nuclear bomb. You need a very well defined model to predict when we reach the critical mass to start a chain reaction. Still, the model is not perfect, as the involved matter is not as pure as the model requires, and also the measurements (energy, atom dimensions) are really measurements approximations. After one year of storage, the mass of the nuclear bomb is not the same anyway. I'm sure there are really big error intervals that are taken into account when applying any "exact" physics prediction. Still, I cannot think about any physics model that don't have as goal, ultimately, to do predictions about reality. Even string theory is a model that ultimately aims to make better and more exact predictions about reality than the relativity theory and quantum theory.
History
History is mainly a backward prediction. We have some observed realities (archaeological objects, stories, documents) and we try to backward predict the string of event that likely created the observed state. We cannot be sure that the documents are not lying, we just do our best to predict the most likely events that would create the history artifacts we can observe now. We cannot go back in time, we just backward-predict. Sometimes we forward predict based on hypothesis about past events, just to eliminate some events that proves to have consequences that do not harmonize with other likely backward predictions about the past. Predictions...
Human knowledge
The human life is a struggle to predict the likely outcome of certain current events, and find the best actions that have the desired outcome. We predict with good reliability who's face we see (while we can sometimes confuse people), we predict where our keys are (while sometimes they are somewhere else), we predict what is the best time to wake up to arrive at work at a reasonable hour.
Everything is forward and backward prediction, trying to optimize some internal objective functions, that are also weak approximations of what is good for us. For example we can do a very good prediction about the optimal way to obtain that smoking cigars that can damage our health. We predict, therefore we exist.
Artificial intelligence
For now artificial intelligence (or "machine learning") is mainly about predicting similarly with what human perception does. Andrew Ng told us a "rule of thumb" that the current "machine learning" is likely to be able to do what the human brain is able to do in less than 1 second, mainly automatic perception.
The computer can also solve a number of real problems that are hard for the human brain, but only when these problems can be formalized in a very specialized model, like in graph theory. The computer can play chess and recently beat even the best human players. Still, the computer can have a hard time to understand some language subtleties. The computer might still be unable to do some cognitive tasks in situations when even a child can do a good job.
The future of machine learning as I see it is better prediction about real life situations. These situations will not be exactly predicted, as the real life cannot be known with accuracy. Even a medical doctor diagnostic is often proven wrong, because we cannot have full knowledge about a patient. In order be become intelligent, the machine should be able to correlate a lot of statistical predictions about the regularities of the world in order to be able to do forward and backward prediction about real life.
This includes predictions about the language regularities, that should be correlated with likely scenario in life. Ambiguities like "the object does not fit in the bag because it is too small" can only be solved by also using some predictions about the behavior of these objects in real life.
Dear reader, please leave a message if you exist! ;) Also, please share this article if you find it interesting. Thank you.
Comments