Or are they fed instructions only to follow?
Machine Education and Computational learning theory employs different models of learning using smart algorithms to make machines intelligent and learn faster. But in theory, do machines actually learn "anything"? They are fed instructions which they follow, whether associative, supervised, reinforcement, or adaptive learning or exploratory data mining, these are all information (instructions) that are fed into computers. There is no involvement of motivation and synthesis of learned materials.
What's the precise definition of machine learning?
Hi Sidharta,
What an excellent question. I am glad to read I am not the only one think in this way.
Unless, I am mistaking machine learning is very often associated with neural networks. The idea behind is to mimic the connection within our brain.
https://en.wikipedia.org/wiki/Machine_learning
You have asked "in theory, do machines actually learn "anything"?". Learning can mean many things, depending you are a teacher, a psychologist, a biologist and the list is not exhaustive. For the sake of simplicity, let's assume that learning modifying existing knowledge then we could perhaps make a link between learning and machine learning.
In my opinion, I prefer comparing machine learning with problem solving. It is true a problem domain with some possible instructions are fed to a "problem-solver" program. Then this program could generate another program to find a solution. This solution needs to be compared against a goal. Newell and Kannt have produced very interesting literature focusing on these issues. At the time, Genetic Programming was perhaps not yet considered.
Since their publications, hyper-heuristics has made an attempt to "learn" some algorithms. I prefer the view that the hyper-heuristics searches the space of solvers. Often many of the algorithms are not human readable or even studied. GP has found some human-competitive design though.
Patricia
Machines can be trained to learn but lack intelligence to learn on their own. This means that they can provide solution under tailor made conditions. This process of learning aims at reducing human time and efforts in problem solving in a given domain.
Very interesting question to ponder about.
In answering this question, I would break the argument into simpler more manageable problems.
1) As Joachim so precisely puts it, learning can be defined as changing the system's state to produce an output(based on the objective function's goal)
2) They have the ability to memorize, cluster, generalize and receive feedback in reinforced learning. All this based on input data given to them, which are common strategies to derive knowledge about the domain under study
What remains to be seen is how machines deal with:
1) Scale up to absorb more information and produce sensible results in multi domain environments
2) Integrate and be able to discriminate among domain differences to produce reasonable output
3) What I find the most challenging baseline is to produce emergent behavior inherent in humans and animals. This relates to higher level abstractions of generalizations from which to generate new knowledge based on what was learnt.
In the last three points, I would answer in the negative at this point in time. The interesting thing is that, in my opinion we are failing on these last three points not based on theoretical limitations, but on the research focus. To be able to test the last three points, we need to focus on merging different algorithms into a complex system. This is not unreasonable to expect, since the brain contains specialized areas and different types of neurons. The other challenge would be in answering if these systems can be leveraged using the existing brain architecture or do we leverage the strength of linear and parallel computing architectures without going into imitating the brain.
Yes machines learn; most effective using machine learning techniques, and gaining the practice by implementing them and getting them to work as machines on their own. More importantly its not only the theoretical underpinnings of learning, but also gain the practical know-how needed to quickly and powerfully apply these techniques to new problems as our world evolve especially in advances of knowledge and expertise know-how...
Believe me machine need to learn even more in dynamic environment compare to human beings , it cannot replace us in the way we learn and adapt. Machine demonstrate based on the ethics provided. Inference make hypothetical answer with multiple outcomes. Right of information cannot be classified until it is primarily authorized. Machine u can learn anything but u still lag instant decision capabilities even with your huge Analytic s. Stay learning !!!!!!!!
Yes they do. Please see the famous "Turing Test" by Alan Turing in the attached Wiki link. Alan Turing, in a 1951 paper, proposed a test called "The Imitation Game" that might finally settle the issue of machine intelligence. The Turing test is a test of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. Alan Turing proposed that a human evaluator would judge natural language conversations between a human and a machine that is designed to generate human-like responses.
https://en.wikipedia.org/wiki/Turing_test
To me "machine learning" is the formation of insight / knowledge from data provided via algorithms & computing power. Machines do learn via supervised learning, unsupervised learning, reinforcement learning etc. Machine learning use cases include self-driving cars, web page & emails classification, search ranking, clustering of customers into different open categories based on similar behaviors etc.
However, as of today the way machine learns is different from the way human learns as machine is not ready to demonstrate the learning from open discussions & lectures, implicit / unconscious learning, learning as a result of various emotions e.g. motivation, curiosity, fear, boredom etc. May be these can be improved in future. If you want can refer to the following link for machine vs human learning:
http://www.cs.cmu.edu/~tom/10601_sp08/slides/HumanMachineLearning_04_2008.pdf
In future, even though if machine can achieve the level of human's learning & cognition ability, but not sure machine can understand & process emotions like human does as human beings have souls (cognition & emotions).
Thanks to all for your much valued comments. What appears from this discussion, and without any dispute is that: machines learn and get trained but not using their own intelligence. Since learning induces change in behaviour, such behavioural change is apparent in human beings more often than in machines, since it is difficult to define learning-induced behavioural modifications in machines. The acquisition of information, and thereof its processing by the brain or its computer equivalent, is a complex cognitive mechanism which enables a system to learn and modify its own behaviour which further augments its auxiliary learning (adaptive learning in machines). In this respect, machines are exposed to dynamic environments as we are in similar sense, and they are programmed to function based on instructional modules (algorithms). Instruction is a common denominator here which is essential for both humans and machines to "learn". Then, what differentiates machine learning from human learning?
No! machines do not learn.
According to [Mitchell Tom. (1997): “Machine Learning”], a machine learning approach attempts to develop strong algorithms that allow machines to improve [the productivity of] their performances on a given goal [and an objective function]. The word ‘learning’ in ‘machine learning’ is ONLY a metaphorical image and is a reflection of human knowledge acquisition and learning in machines and artificial agents.
Dear Farshad, your comment is interesting and this is about what we are discussing here. Convincing evidence supports the fact that machines, for instance, if you consider robots, can be taught to perform several tasks based on given instructions. Better algorithms are developed to improve on their functions. Machines do not perceive knowledge in a way we do, even if the most advanced robots using smart algorithms can "see" and read given instructions. I was not aware of Tom Mitchell's work, and would love to read how he explained machine learning. Well I forgot to ask, do they build 'concepts'? I guess they do not.
Regards, Sidharta
Thanks for the link Joachim. As far as one can comprehend, this paper addresses and introduces some novel pattern recognition techniques in order to enable machines to read and interpret handwritten characters. Of course this is a mode of machine learning that involves teaching machines how to read characters and emulate human learning, but the scientists also acknowledge the fact that machines require far more information to recognize simple characters compared to humans who require a very few of them, since humans can build concepts and relate those to mental constructs. One aspect that they mention which is very interesting - the machine 'learns to learn' employing adaptive algorithms that are smart enough to exploit the given data. Note that machines are extremely good at emulation and data mining, they are fast and efficient, yet they lack power of abstraction.
Yes Machines can learn, but no, they don't understand what they have learned.
Dear Graeme, that's well said. Perhaps that somehow exemplifies the definition of rote learning, although machine learning do not involve such repetition, yet they are able to recall information with precision. There are many factors that differentiates machine learning from human learning, however, as Joachim above indicates that this is due to differences in technical realization, it may be compelling to assume that to some extent this could be overcome when such differences are reduced. Farshad has pointed out to an interesting reference/book which contends that machines do not learn, they just memorize. In fact memorization and learning are very much connected to each other, since it is necessary to learn in order to memorize, but not the other way around (not necessary to memorize what you have learned). Associative learning /relational learning do not require memorization to such great extent (but memory is still relevant) wherein learned facts could be linked or associated to some constructs to build concepts in order to recall them whenever required. Anyway, this is interesting.
Dear Sidharta,
I do agree. And as you mentioned, the most important focus is on 'concepts'. Concept is a very complicated term that must be used with caution. Based on a very general description concepts are the products of the human's mental images that manifest themselves in languages and expressions. Accordingly human transforms them into hypotheses in order to train them to the machine's knowledge bases. So, the machines do not construct concepts, but they expand their hypotheses (generally based on inductive rules). This has been known as Class Expression. Therefore, in my opinion machine learning is nothing but hypothesis generation and reasoning based upon generated hypotheses.
Yes Sidharta, there is an architectural difference between the way the brain "Understands" it's knowledge and the way that Machines such as the computer are built. The main difference is the Massively parallel Sub-layer that does hierarchical Analysis on the data at the time of, or just before, the serial data is generated. Because we only recognize the serial data, in Computers, the massively parallel data never gets created.
In the debate whether machines learn, I like to put forth the question in a different light. How does "human" learning differs from the already established statistical learning theory? From my perspective, there has not been a logically consistent theory that invokes a type of learning that goes beyond what statistical learning theory already covers besides complexity of the system. The ones that do can be solved by a machine through heuristics.
Fuzzy abstract concepts that are usually invoked to criticize AI are also problematic to human beings and have been debated by philosophers throughout history (and there is no end to the debate in sight). It might be argued that we still can deal with highly debated concepts, yet these are also subjectively applied rules which are learned from the environment. Also , to make abstract concepts you can put neural networks to generalize upon a system of neural networks that have diverse training. How would the concepts derived from this system diverge from our biological neural network?
The other concepts put forth by philosophers such as the Chinese room argument are also problematic. The problem being that learning and understanding are not defined mathematically and reduced to quantifiable experiments(having humans and machines in the same footing without invoking untestable elements of human nature) that can disprove the capability of machines and can fit in a consistent theory(if there were, we would not be having this discussion).
I still think the AI community has gone astray and I am highly critical of the research directions taken(I still have hopes for AGI). Nonetheless, from my perspective there is no fundamental element in my view that limits in any way machines from learning to the capacity of the human being.
Just my biased opinion
At the start of the discussion, it was suggested to define the notion of learning, before we can consider whether or not a machine can learn. I believe machine learning is just an automated solving machine.
Nonetheless, this article will show how our different cognition process have been modelled in a pyramid. They are common elements, but some of them would be hard to implement in a computer system. And if we do, perhaps it is the rise of the machine.
Patricia,
Thanks for the excellent reference and it is a good starting point to then place the current conversation in a tangible fashion on which to question machine learning.
All,
To highlight some of the minutia of semantics and loose reasoning that often lead to this discussion to head in diverging paths, take the reference given by Patricia on page 4 it states "It is noteworthy that unlike a machine, human abilities for problem solving do change depending on ages and external influences". Again I could argue for a system that has coded the different algorithms on Duda's book and use genetic algorithms to select a series of steps using the algorithms and based on an objective function,analyze the results and choose the best algorithm mix. Wouldn't that equate to a student sitting on a calculus course memorizing the formulas and choosing the best formula for the given problem?
A counterpoint would be on two fronts:
a) The machine did not derive the algorithm. The response would be that the student also did not and do not have to do the derivation to solve the problem.
b) The machine did not understand the problem and just used the algorithm provided. I would highlight the author's evaluation of Wallas and Polya work on page 5 as based on "introspection and informal observations" by using terms such as "Understanding the problem" and "devising a plan".
While challenging it is not insurmountable to construct a system that can fulfill the notions provided in the paper as long as the implementation is not limited to toy implementation algorithms as portrayed in the literature and true systems are used .
I have no quarrel in revising my posture, once we have a strong mathematical formulation of the problem that judges machines and also the human ability on equal footing without reverting to notions such as those of Wallas and Polya
Yes, I think the human learning is somehow the same; fed by instructions from the world and gradually becomes more and more accurate to deal with things. Machine Learning is much more simplified than what is happening to man to learn. The instructions are values so the machine would be able to use them, and the methods are much simpler that what's happening in the process taking place in human brain.
Dear all RG members and Colleagues,
Wish you all as Merry Christmas and a very Happy New Year 2016!!!
Best Wishes, Sidharta
- Badges
- Science topic