User:CyborgDan

Science Review
This is a work in progress review of a very interesting area of computer science called machine learning.

version 2.2 Added more references. version 2.1 Final touch ups. version 2.0 Review Structure enforced. Major work over of article re-enforce ideas through cited texts. References Added. version 1.1 Ideas collated in to coherent paragraphs. Version 1.1b Initial Thoughts no cited texts yet. A mere brain storm of ideas to convey, structured in to categories. Version 1.0 Initial Document.

Learning Computers For Dummies
Intro

Machine learning is a hot topic in the world of computer science. It is a form of artificial intelligence that is entirely centered around the learning of a machine. The ultimate goal of machine learning is to gain full independent learning. This review is intended to give a clear review of the major aspects of machine learning. It touches on major issues with machine learning and current mile stones.

Common Misconceptions and Build Basis

There are a number of misconceptions of computer intelligence and machine learning is that an intelligent system can learn. Machine Learning pertains to a computer being able to actively acquire knowledge with out the aid of a human or some other learning agent. However a machine can be still be intelligent, but unable to learn on its own, and these machines fall into an area of study called intelligent systems. A learning machine is able to build its own reasoning through the experiences it has learned, while an intelligent system is implanted with all the logic and reasoning need. So as to gain a basis for this review, machine learning shall be solely related to a machines ability to independently learn and acquire valid reasoning.

The main elements of learning include information acquisition and reasoning, while machine learning at its core is very mathematical so it seems fair that the learning process be confined to the main elements of human learning, which include deductive and inductive reasoning. Natural Human learning is deductive, as a person is presented with a stimulus they unconsciously deduce the closest match to the given stimulant through logical connections that have been established by past experiences. Inductive reasoning is the opposite process where a person can deduce the original state of a stimulus knowing the outcome. These simple elements provide a solid basis for machine learning, as computers can only operate within strict concise instructions.

Human Learning and Machine Learning

Their are many different ways of learning for any human being, and the same goes for machines. In fact there are a number of different proven ways for a machine to learn, through a set number of constraints. However for simplicity this review will concentrate on form of computer learning called reinforcement learning. Reinforcement learning comes from the premise that a person should be rewarded for a step in the right direction, so as to reinforce this decision over others. From a pure machine perspective with no prior knowledge and the machine contains “inductive” logical reasoning. The machine will apply reinforcement learning to all possible solutions to any problem presented, and each solution will be given a weight that corresponds to how close to a solution it is. While this is a valid and proven method of learning, it does not allow for random variations (“noise”) to affect the learning process. Noise pertains to random variables that effect the learning process, and these can have a positive or negative effect on reinforcement learning as a machine might have to relearn a process it has already deduced.

Algorithms and Hardware

Current machines operate in a repetitive state, with little to no abstraction. Abstraction is the process of grouping data into groups based on characteristics. For clarity humans abstract data based on natural language, so everything we learn is abstracted in our mind to natural language with which we easily understand and can apply logical reasoning. While machines operate in mathematical state space which is very strict and allows only true or false. If a computer was to be abstracted to allow fuzzy logic (Half Truths “mostly true”, “mostly false”, etc) the process of reinforcement learning, presented earlier, would be dramatically slowed because the possible number of solutions increases at an exponential rate. However if technology was to catch up to the advancements made so far in machine learning its theoretically possible to have a computer that is capable of full independent learning, with some level of abstraction.

By far the main concern with machine learning is not the learning process, but merely the countless storage space required to store the massive amounts of data that is acquired. There are many factors that lead to massive memory usage, most are linked to the ways information is feed to the machine. Take the fuzzy logic abstraction example presented above, if data was streamed to the machine uncompressed, it firsts needs to be processed and stored, however the computer has not learned to remove unwanted data “noise” from the information it is storing as such memory usage will rapidly expand. A machine is not aware of how much memory it has, as such it can never learn to remove “noise”. It is the limited memory and machine knowledge of itself that has presented a big problem to machine learning.

Conclusion

Machines cannot compete with humans, in the ability to learn the simplest of tasks