New machine learning theory raises questions about nature of science
By: Princeton Plasma Physics Laboratory
So what is science? This machine learning algorithm doesn't utilize a body of scientific knowledge and doesn't utilize the scientific method.
The algorithms empirically arrives at predictions using only observations. While it may be possible to argue that the algorithms are empirically and independently developing the laws of physics to make predictions, the machine does not generate those laws as a result.
S E E D E D C O N T E N T
A novel computer algorithm, or set of rules, that accurately predicts the orbits of planets in the solar system could be adapted to better predict and control the behavior of the plasma that fuels fusion facilities designed to harvest on Earth the fusion energy that powers the sun and stars.
The algorithm, devised by a scientist at the U.S. Department of Energy's (DOE) Princeton Plasma Physics Laboratory (PPPL), applies machine learning, the form of artificial intelligence (AI) that learns from experience, to develop the predictions. "Usually in physics, you make observations, create a theory based on those observations, and then use that theory to predict new observations," said PPPL physicist Hong Qin, author of a paper detailing the concept in Scientific Reports. "What I'm doing is replacing this process with a type of black box that can produce accurate predictions without using a traditional theory or law."
Qin (pronounced Chin) created a computer program into which he fed data from past observations of the orbits of Mercury, Venus, Earth, Mars, Jupiter, and the dwarf planet Ceres. This program, along with an additional program known as a 'serving algorithm,' then made accurate predictions of the orbits of other planets in the solar system without using Newton's laws of motion and gravitation. "Essentially, I bypassed all the fundamental ingredients of physics. I go directly from data to data," Qin said. "There is no law of physics in the middle."
The program does not happen upon accurate predictions by accident. "Hong taught the program the underlying principle used by nature to determine the dynamics of any physical system," said Joshua Burby, a physicist at the DOE's Los Alamos National Laboratory who earned his Ph.D. at Princeton under Qin's mentorship. "The payoff is that the network learns the laws of planetary motion after witnessing very few training examples. In other words, his code really 'learns' the laws of physics."
Machine learning is what makes computer programs like Google Translate possible. Google Translate sifts through a vast amount of information to determine how frequently one word in one language has been translated into a word in the other language. In this way, the program can make an accurate translation without actually learning either language.
The process also appears in philosophical thought experiments like John Searle's Chinese Room. In that scenario, a person who did not know Chinese could nevertheless 'translate' a Chinese sentence into English or any other language by using a set of instructions, or rules, that would substitute for understanding. The thought experiment raises questions about what, at root, it means to understand anything at all, and whether understanding implies that something else is happening in the mind besides following rules.
Qin was inspired in part by Oxford philosopher Nick Bostrom's philosophical thought experiment that the universe is a computer simulation. If that were true, then fundamental physical laws should reveal that the universe consists of individual chunks of space-time, like pixels in a video game. "If we live in a simulation, our world has to be discrete," Qin said. The black box technique Qin devised does not require that physicists believe the simulation conjecture literally, though it builds on this idea to create a program that makes accurate physical predictions.
The resulting pixelated view of the world, akin to what is portrayed in the movie The Matrix, is known as a discrete field theory, which views the universe as composed of individual bits and differs from the theories that people normally create. While scientists typically devise overarching concepts of how the physical world behaves, computers just assemble a collection of data points.
Qin and Eric Palmerduca, a graduate student in the Princeton University Program in Plasma Physics, are now developing ways to use discrete field theories to predict the behavior of particles of plasma in fusion experiments conducted by scientists around the world. The most widely used fusion facilities are doughnut-shaped tokamaks that confine the plasma in powerful magnetic fields.
Fusion, the power that drives the sun and stars, combines light elements in the form of plasma—the hot, charged state of matter composed of free electrons and atomic nuclei that represents 99% of the visible universe—to generate massive amounts of energy. Scientists are seeking to replicate fusion on Earth for a virtually inexhaustible supply of power to generate electricity.
"In a magnetic fusion device, the dynamics of plasmas are complex and multi-scale, and the effective governing laws or computational models for a particular physical process that we are interested in are not always clear," Qin said. "In these scenarios, we can apply the machine learning technique that I developed to create a discrete field theory and then apply this discrete field theory to understand and predict new experimental observations."
This process opens up questions about the nature of science itself. Don't scientists want to develop physics theories that explain the world, instead of simply amassing data? Aren't theories fundamental to physics and necessary to explain and understand phenomena?
"I would argue that the ultimate goal of any scientist is prediction," Qin said. "You might not necessarily need a law. For example, if I can perfectly predict a planetary orbit, I don't need to know Newton's laws of gravitation and motion. You could argue that by doing so you would understand less than if you knew Newton's laws. In a sense, that is correct. But from a practical point of view, making accurate predictions is not doing anything less."
Machine learning could also open up possibilities for more research. "It significantly broadens the scope of problems that you can tackle because all you need to get going is data," Palmerduca said.
The technique could also lead to the development of a traditional physical theory. "While in some sense this method precludes the need of such a theory, it can also be viewed as a path toward one," Palmerduca said. "When you're trying to deduce a theory, you'd like to have as much data at your disposal as possible. If you're given some data, you can use machine learning to fill in gaps in that data or otherwise expand the data set."
Tags
Who is online
102 visitors
Empirical correlative results can be just as accurate as anything generated by the scientific method. And there isn't any need to know and understand physics.
So what is science?
Sound to me simplifies as:
See where its been, see where it is now and follow the same trajectory forward.
The machine did not utilize physics to obtain results. Is it physics if physics isn't required?
The machine did not utilize the scientific method or scientific laws. Is it science if science isn't required?
Is it natural to look to the past, see the present and be able to anticipate the future ?
To me it is.
But, If a computer does that is that ai ?
Considering that physics and science developed the machine. Where would it be without them ?
OMG Nerm.... this again?
People don't seem to realize that machine learning is creating existential questions. In the case of the seed, what is science?
The machine doesn't need physics and, by extension, doesn't need physicists. The machine only needs data.
Training scientists is a long and expensive process. The machine can be built quickly and will train itself. Will machines replace scientists?
It's doing no such thing. It's not a true AI. It is doing what it's programed to do. Just like any other machine.
That's like asking if machines will replace humans.
Doubtful. Humans will to survive is too embedded in our genes.
They/we will fight the takeover by ai.
Remember Ted Kaczynski ? That was his fear and his irrational fight, Artificial intelligence taking over humanity.
It is not possible to make a LASTING compromise between technology and freedom, because technology is by far the more powerful social force and continually encroaches on freedom through REPEATED compromises.
Human's will to survive, is in artificial intelligence's way.
Machines are a tool that humans use, Machines using humans is a nightmare few desire.
The machines will do what humans are currently doing. That suggests machines will replace present human function and purpose.
Human function and purpose may be relegated to the role of eating, excreting, and reproducing.
That's quite the imaginative stretch. This isn't the age of the Jetsons or something like that. Machines are just tools we use.
I've said for years IF I were a young person I'd be studying robotics and artificial intelligence. Yes it is the future.
But the human will to survive will endure. It's "programmed" into us all.
But what will be the role of humans in the future? If humans become dependent upon machines for survival then who is controlling who?
Many many humans survival now already depends a great deal on machines in many ways. I have no doubt that will increase.
Unfortunately even today people will die because of machine failure as they freeze to death from lack of heat due to no electricity.
who is controlling who?
For now, humans are. What controls us in the future comes NEXT.
lol .. pun intended
Machine learning is, at its core, pattern recognition. Basically it is a way to approximate an arbitrarily complex function by learning patterns from mountains of data.
A machine learning algorithm itself is not science. It can be, however, a theory. Here is how:
The hypothesis is that the source data (that which the algorithm processes) contains sufficient information to infer the rules of physics (for the application in question) to a satisfactory level of accuracy.
The algorithm (once trained) produces results that can be compared with empirical evidence to see how well they correlate.
If the results correlate sufficiently well with empirical observations, the machine learning algorithm is indeed an expression of a theory. It is, if you will, a theory expressed as a computer function.
That's the question isn't it? Are the algorithms actually inferring the rules of physics?
As I understand it, the algorithms are provided orbital data that conforms to patterns, identifies correlative relationships between those orbital patterns, and predicts orbital patterns of planets for which data was not provided. The article doesn't indicate what operating conditions were used to initialize the process, such as the number of planets to include in the prediction.
The algorithms are not provided the rules of physics and do not produce the rules of physics as a result. The algorithms only make predictions but doesn't explain how anything works. The algorithms aren't really using the scientific method.
So, is the information generated by the algorithms a scientific result? Or is the information an empirical result that has little if anything in common with science other than arriving at the same result?
Yes! They are approximating the rules of physics within the constrained area of study.
Correct, the algorithms are provided empirical data. Why do you simply claim they do not approximate the rules of physics? They do, that is precisely the kind of thing machine learning algorithms produce. Think of it as extremely complex, multi-dimensional curve fitting because that is ultimately what the underlying mathematics produces.
Quantum theory does not explain how particle physics works, it simply makes predictions. Extremely accurate ones too ... the best in science.
Seems to me that this is what you wish were true and are going to simply keep repeating this as a mantra.
As before, I think you just want this to not be deemed a 'scientific result'. But the facts are this. Empirical data is processed by a complex algorithm which is able to make predictions based on a multidimensional curve fitting function inferred from the data. Based on empirical data, the algorithm found patterns, formulated a function and can now execute that function to make accurate predictions.
Because I do not know if the algorithms do approximate the rules of physics. Obviously the algorithms are not applying the rules of physics. The algorithms produce the same results as do calculations using the rules of physics but that, alone, doesn't equate to inferring the rules of physics.
The rules of physics can be used to make predictions. But that is because the rules of physics describes causal relationships. The algorithms are making predictions without describing causal relationships. So are the predictions made by the algorithms a scientific result? Has the algorithms performed science to make predictions?
Quantum theory is not applied using an empirical process.
Where does the scientific method fit into the process you describe? If the rules of physics are unnecessary for producing accurate results then where does this fit into descriptions of science?
What then do you think causes the accurate predictions? The 'rules of physics' are nothing more than manually derived patterns. Here we have a machine deriving the patterns.
Are you playing a semantics game? Nerm, machine learning algorithms do not produce formulas like those human beings use. They do not recreate the relevant portions of Einsteins ToR. If you are trying to argue that they do not derive the formulas that human beings used (e.g. Einstein's field equations) then yes of course that is true. The algorithms have an entirely different representation for the multidimensional function.
The results though, as claimed by the seed, match the results produced by applying the manually derived rules. Different representation, but the functional results are the same. That means the algorithm has successfully derived the relevant physics of the problem domain. The underlying relationships of the algorithm, however, is no doubt a complex neural network that has no comparison in form to the manually-derived formulas taught in universities.
A machine learning algorithm will, once trained, accept a multidimensional input condition and produce a prediction (which may itself by multidimensional). It is a prediction just like calculating using the mathematical equations of a theory is a prediction.
Evidence ... Pattern recognition (hypothesis) ... validation .... theory ... repeated prediction and verification of theory.
No, I'm not playing semantics. The machine learning algorithms do not have the rules of physics available. What the machine is doing would appear to be more akin to the work of Copernicus and Kepler than that of Newton. Copernicus and Kepler did not have the rules of physics available and did not employ the scientific method. Copernicus and Kepler were grounded in the Aristotelian approach to natural philosophy.
The machine learning algorithms are employing empirical logic based on mathematics. The machine isn't theorizing, hypothesizing, or asking questions. The algorithms are essentially reverse engineering a mechanism in a manner similar to that of Copernicus and Kepler. No rules of physics required. The machine learning algorithms doesn't need Newton, Einstein, or quantum physics any more than did Copernicus and Kepler.
Is it physics if the rules of physics are not required? Is it science if the scientific method and rules of science are not required?
What is science?
Yes, Nerm, I have repeatedly affirmed that. It is fundamental to the concept of machine learning to infer the rules from the data.
It just never sinks in. I gave you quantum physics as an example on purpose. Science does not understand how particles behave / manifest as they do, but quantum mechanics predicts the behavior with uncanny accuracy. The rules at the particle level remain to be discovered. You would argue that the field of quantum physics is not science.
The algorithm produces a theory based on observations that is predictive and falsifiable.
I am done repeating myself.
Your 'is it science' mantra seems like a stubborn, set-in-ways slide-rule engineer of the 1950s who rejects 'electronic calculators' and insists that all calculations must be done by hand to be 'true'.
What exactly is your issue here Nerm? TiG has repeatedly explained the concept to you. So what are you still confused about? This just seems to be another rant against science (i.e. is it/what is science) in a series of long running rants against science from you. You pulled the same BS against the science behind Covid precautions.
No, that isn't what's happening. The machine algorithms produce a prediction; not a theory.
People playing the lottery are making predictions that are tested and are falsifiable. That only requires choosing numbers, so predictions do not require theories. People do attempt to identify patterns from past lottery results but the lottery isn't mechanistic in nature as is the solar system. Are people's attempt to identify patterns in the lottery data science?
The machine learning algorithms are not making predictions based upon physics. The predictions are derived from repeating patterns in the data. The machine doesn't care if the data represents gears or planets. If the machine is identifying causal rules then it is the rule of a repeating pattern and nothing more.
What is science? More importantly, can machines replace scientists? Will science become nothing more than data processing? Will there be a need for an Einstein in the future?
My issue is that TiG isn't addressing the question 'what is science'?
The machine doesn't need physics and, by extension, doesn't need physicists. According to what TiG has presented, machines can do the work of scientists.
The seeded article points to the possibility of using machine learning in the development of fusion reactors. Based on the planetary model of machine learning, the machine will not require science or scientists. The machine will only require data. Will that be science? Or will that be data processing?
Is science nothing more than data processing that can be performed by machines?
Science is a systematic process. That's been explained to you before. So I'm not sure why you ask the question or what you expect? Data collection and usage is simply a part of that process. It doesn't matter if it's humans or machines doing it. It seems you dislike the idea of machines being a part of the process?
The machine is recognizing repeating patterns in the data. Will correlation replace causation? What impact will that have on future science?
The machine makes predictions based on the data. That's it. You're reading way too much into it otherwise.
Which is what I've said. Are predictions based on the data science? If that is science then obviously machines can replace scientists, right?
I'm not the one who made esoteric arguments that the machine is performing science. I'm not the one who argued that the machine is inferring the rules of physics. I've pointed out that the machine is only processing data to identify repeating patterns and to make predictions based upon learned recognition of repeating patterns.
If the machine is performing science then why are scientists necessary? Why is scientific knowledge necessary? The machine doesn't need to be provided the rules of physics to make accurate predictions.
Your questions have already been answered by TiG. I see no reason to repeat him or myself when the answers have been given. So at this point, you either do not comprehend what the machine does or are being intentionally obtuse about it. Which is it?
I guess I'm being intentionally obtuse because I don't accept that the machine is performing science. I see TiG's argument as being demonstrably incorrect and I disagree with TiG's arguments.
Bluntly, TiG's answers are incomplete and incorrect. Machine learning is not confined to the pedantic limitations that TiG is attempting to impose.
But there is a conundrum. If the machine is not performing science then what of Johannes Kepler?
A scientific theory is a mechanism that, based on perceived patterns in observations, can accurately predict future observations within its problem space.
Again, you would deny quantum theory because that is all it does. Quantum theory is a black box which predicts but does not explain. Same here; the algorithm has inferred patterns in observations and can now accurately predict future observations within its problem space.
Empirical observations → empirical theory → predictions
Bluntly, I have and currently am building a machine learning algorithm. Have you? Do you have expertise in this area? I see no evidence of it.
Further, what limitations have I imposed on machine learning?? More made up crap, Nerm.
Scientists represent their findings in mathematics. The mathematics are then used by other human beings (e.g. engineers). Machine learning can produce theories and thus can indeed operate in the role of scientist in that regard. But it cannot build the human mathematics (i.e. it cannot ... currently ... build Einstein's field equations). A cyber scientist does not replace scientists and is not a replacement for science itself. It is, however, part of contemporary science. It is an alternate means to produce empirical theories. Science has evolved and machine learning is part of it.
It is most definitely possible that machine learning (automatons) can perform far more advanced functions of science. I am expecting that, in the future, we will have sophisticated automated researchers who will make breakthroughs in areas like cancer research. These automatons (cyber scientists), unlike their human counterparts, will (as is true today) process unfathomably large volumes of data and detect multivariate patterns in the data far better than human beings. As time marches on, their capabilities grow geometrically.
Even so, this will not replace science and scientists in our lifetimes. To replace the creativity and imagination of the human mind is an idealized goal of AI. While this might be possible in the future, none of us will live long enough to see it.
This is the kind of crap that encourages me to dismiss everything you write. You do not accept that the algorithm is performing science (seeing it as lesser than science) and in the next sentence complain that I am imposing limitations on the algorithm.
Yes, I suppose you are. But you are not in a position to dismiss TiG's explanations, much less declare them incorrect, as you have no expertise or credibility to legitimately do so. Especially since you offer no logical rebuttal. Your disagreement with him is irrelevant and largely dismissed, as your comments indicate a lack of comprehension (willful or otherwise) of the machine issues and TiG's explanations.
Where does the scientific method fit into that process?
Doesn't that process describe Aristotelian natural philosophy? And that Aristotelian process places greater emphasis on correlation rather than causation. If the predictions derived from correlation prove accurate then why is an understanding of causation needed?
Good for you. Back when I was attempting to do this kind of stuff the equipment was too limited. It was necessary to use a hybrid approach where the learning involved both data processing by machine and interpretation by scientists.
8088, PDP-11, and DEC VAX computers simply didn't have the computational speed or data storage to perform machine learning. Today's desk top computers have vastly greater capabilities than did main frames back then. Yes, I have been involved in creating hybrid learning systems that mixed machine processing with human involvement in the decision tree. That was the frontier of available technology at the time.
And now machines can accomplish the same thing without the need for scientists or the need to provide the machine with scientific knowledge. The machine learning algorithms highlighted in the seed article has removed the scientific method and scientific knowledge from the process.
If a machine can make accurate predictions without being provided the rules of physics (without being provided basic scientific knowledge) then what is the machine doing? Claiming that the machine is independently developing the underlying rules of physics without prior knowledge raises questions about the nature of science.
But that does not address the existential question of what makes science science. What differentiates science from other activities that process data to arrive at predictions? Is a sales manager a scientist? Is a production line foreman a scientist? They both scrutinize, process, and utilize data to make predictions.
Machine learning may be doing something other than science. That doesn't invalidate machine learning. But that avoids warping science into something it's not just to accommodate machine learning. So, what is science? And is machine learning really performing science? Acknowledging that machine learning is doing something other than science doesn't invalidate machine learning.
Asked, answered, repeatedly. TiG@2 @2.1.1 @2.1.3 @2.1.5 @2.1.16
No, it perhaps means that some of the math done by scientists can be done by the computer. This in no way changes the laws of physics and what they mean for us as humans. This AI is simply doing the calculations that human intelligence has previously done.
Not really, the calculations fed into the machine as he said "he fed data from past observations", the only difference is that the human calculated predictions weren't fed in and the machine came up with the same answers humans already knew. It's not like it's an AI that is gathering data on its own and coming to its own calculations and conclusions.
Math. It's still being provided "data from past observations", it's just doing the calculations it was programed to do by a scientists using the laws of physics.
It isn't doing it "without prior knowledge", it's simply doing it without certain prior knowledge but with the same data that scientists used to understand and define the laws of physics.
Nothing presented in the seed puts science into question. The methods used are still sound, the conclusions still reliable.
Virtually all humans have been learning by trial and error for eons, and no, just because one employs some aspects of observational data crunching to form a conclusion, it does not make one a scientists no matter how much some young earth creationists wish it. In a closed system where the sales manager only has data inputs from a narrow source, they may come to the conclusion that life is empty and worthless. Simply crunching the numbers and coming to individual conclusions does not mean you're a scientist, just a layman imitating singular aspects of a scientists, like a car owner changing his oil isn't actually an auto mechanic.
If you have to ask that then I suggest you do your own research as science is a vast collection of the studies of the natural universe.
Not any more than a calculator is performing science. Can it be an invaluable tool for humans to better understand and draw conclusions? Sure, but it's not actually "doing science" since this machine would come to wildly different conclusions if the previously observed data came from porn hub instead of Hubble.
Exactly. The machine is a computation engine and nothing more. The most basic type of machine learning is to feed data into a spreadsheet, perform a statistical analysis to generate a polynomial equation, and make predictions using that polynomial equation. The entire process can be automated. But the spreadsheet is only performing calculations and identifying correlations; it's not independently developing the rules of physics.
It's a rhetorical question used quite often for philosophical examination and evaluation of situations.
I can assign a lab technician the task of performing an experiment. The lab technician is performing an an assigned task. Does performing the assigned task for scientific purposes mean the lab technician is a scientist? Does performing the assigned task mean the lab technician is performing science? What differentiates a lab technician from a scientist? Answering those questions requires an understanding of what makes science science. And that understanding requires answering the question 'what is science?'
No one ever said it did. You're the one who inferred that.
Philosophy is a separate field from science.
Not necessarily, to be considered a scientist one must have studied a specified field and understand the previous conclusions previous scientists have formed based on their experimentation so that the new scientist has a base of scientific knowledge on which to build and thus does not have to repeat all previous experiments on their own. However, if one wanted to be general about the meaning of scientist anyone who systematically gathers and uses research and evidence, to make hypotheses and test them, to gain and share understanding and knowledge they could be considered a scientist even without a degree.
Yes, if you are taking part in a scientific experiment you are by definition "performing science" whether you're an accredited scientist or not.
That question has already been well defined, attempting to tip toe around the edges of it so as to make it seem less concrete is a pointless activity and exposes one as having some anti-science agenda by attempting to subvert known scientific conclusions or the scientific method itself often due to that persons baseless conflict with those scientific conclusions.
Science: noun - the intellectual and practical activity encompassing the systematic study of the structure and behavior of the physical and natural world through observation and experiment.
Trying to use philosophy, which is just a theoretical exorcise which examines a particular branch of knowledge or experience, to redefine science can be fun for those who want to falsely imagine their own conclusions that conflict with science as possibly correct even though there's no evidence to support them, but it is an exercise in futility because all you'll ever end up with are theoretical conclusions which, by their nature, are just wishful thinking and fantasy without evidence. You might as well spend your time theorizing who's stronger, Hulk or Superman.
The machine isn't observing anything and isn't performing experiments. The machine doesn't differentiate between data obtained from nature and data obtained from man-made systems. The machine would process traffic data or financial data in the same manner as it is processing orbital data.
The machine is only performing an assigned task. The machine isn't performing science.
Yes, a theoretical exercise (a thought experiment) used to scrutinize and test the real world. The point is that the machine learning algorithms are being presented in a way that changes the meaning of science. The machine is performing an assigned task that scientists have performed in the past. There is an inductive leap of logic being made to claim the machine is performing science without the need to observe, experiment, or utilize scientific knowledge.
The machine made accurate predictions. And the inductive leap of logic is that since physicists were required to make those predictions in the past then it follows the machine is learning, inferring, approximating, and independently developing the rules of physics to make predictions. That's a leap of logic that raises questions about what makes physics physics.
Is physics the study of well ordered, repeating patterns found in data?
Exactly, so why are you asking questions you already know the answer to?
No, as you already pointed out, the machine isn't performing science, it is performing a task assigned to it by a scientist. That in no way changes the meaning of science.
I don't see that at all, one can't claim there is "no need to observe, experiment or utilize scientific knowledge" when they are directly inputting "data from past observations".
Based on "data from past observations".
No, no one need make such an inductive leap. Just because you can input a series of pattern data into a painting computer and then let it run on random and it comes up with computer generated "art" doesn't mean computers will replace artists. Just because you can input past observational data into a machine and it will spit out a prediction based on that data does not mean physicists will be replaced, more likely the physicists will use it as a tool like every other machine programmed to make some of the minutia of a job easier. A physicist is able to see correlation across numerous platforms that the machine will never have access to and thus the physicist will always be necessary if we truly want to explore the underpinnings of the universe.
Superman is stronger than Hulk. Supes is also more powerful than Goku. Just saying.
Silicon Valley pioneer Paul LeBlanc joins forces with Special Agent Shea Salazar to stop the A.I. he created.
Scary premise, cool show though