| |
Learning is not an easy word to define since it includes various processes. If you picked up the dictionary and looked for the meaning of learning, you would come across phrases like, "modification of a behavioral tendency by experience," and "to gain knowledge, or understanding of, or skill in, by study, instruction, or experience.” This book will take a look at the different processes of machine learning in a way that is similar to how psychologists study learning in human beings. The spheres of machine learning and animal learning are interlinked. Some techniques used in machine learning are often derived from techniques used in human learning. Some breakthroughs in machine learning can bring to light some facets of human or biological learning.
When looked at from a higher level, it can be stated that a change made to the structure of a machine, concerning changes in the memory or composition, to enhance the performance of the machine, is a sign of learning in a machine. When machine learning is studied at a deeper level, only some changes made to the structure are considered as learning for a machine. For instance, consider a machine that is meant to predict weather forecasts in a certain area for a few weeks. The system is fed with information on the weather over the past year. The system uses this information to predict the data accurately for that city or state. This is an instance of machine learning.
Machine learning is a field that can be applied to machines associated with artificial intelligence. These machines are often responsible for tasks such as recognition, prediction, diagnosis and other similar tasks. These machines learn from the data that is fed to them. Since this data is used to train the machine to perform its functions, it is called training data. Machines “learn” from the training data and continue to learn when new data is fed to them. Machines learn to analyze patterns within the data and use those patterns to address the problem at hand. Different learning mechanisms are used to analyze the training data based on the problems that need to be solved. These mechanisms can be classified into three – supervised learning, unsupervised learning and reinforcement learning.
Skeptics of the field often wonder why a machine would ever need to learn. They may also wonder why machines cannot be produced to address a specific problem alone – like a tractor or a crane. There are many reasons why machine learning is advantageous. It has been mentioned earlier that any advancements made in machine learning will help people understand human learning better. Machine learning also helps to improve the efficiency and accuracy of machines. Some other reasons are:
• Some tasks cannot always be defined even if the greatest engineers were to work on those tasks. These tasks will need to be explained to the machine using examples. The aim is to train a machine to produce output when some input is fed to it. This way, the machine will know how to deal with future inputs and process them to reach appropriate outputs.
• The fields of data mining and machine learning are intertwined. Data mining is a science that deals with identifying the relationships and correlations that lie within and between instances in massive volumes of data. This is another advantage of machine learning in that it might lead to the finding of important information.
• On many occasions, it is possible that humans design machines without correctly estimating the conditions in which they will be functioning. Surrounding conditions can play a huge role in the performance of the machine. In such cases, machine learning can help in the acclimation the machine to its environment so that the performance is not hindered. It is also possible that environmental changes might occur and machine learning will help the machine to adapt to these changes without losing out on performance.
• It is a known fact that change is constant and this is true when it comes to technology. There are changes that occur in vocabulary as well – newer words are added to the dictionary making it essential for a machine to be redesigned. However, redesigning a machine every time there is a change is not practical. Instead, machine-learning methods can be used to train the machines to adapt to these changes.
• Another loophole in the process of human beings hardcoding the process into the machine is that the process might be extremely elaborate. In such a case, the programmer might miss out on a few details since it would be a very tedious job to encode all the details. So, it is much more desirable to allow the machine to learn such processes.
Machine learning is an intersection of different statistical, mathematical and psychological subjects. Each subject mentioned in this section brings a new methodology that can be incorporated in both machine learning and artificial intelligence. These concepts together form machine learning.
Statistics is one of the most common subjects used in machine learning. One problem with statistics that are used in machine learning is training. Training is the process where samples of data drawn from a population are used to identify the properties of the population and also predict the properties of a new sample that is drawn. Another problem is to estimate the value of some functions at a certain point based on values of other functions. Solutions to such problems are instances of machine learning, since problems that involve the estimation of future events often use data about past events. Statistics form an extremely important part of machine learning.
The concept of neural networks, a section of brain modeling, is closely related to machine learning. Scientists have suggested that one possible model for neurons or the neural network is non-linear elements with weighted inputs. Numerous studies have been conducted on these non-linear elements in the past decade. Scientists working in the brain modeling sector are now interested in obtaining information on human learning through the study of neural networks. Sub-symbolic processing, connectionism and brain style computation are a few spheres that are associated with these types of studies.
The control of systems is studied under a subject called control theory. One problem that every system faces is the change in the surrounding environment. Adaptive control theory is a part of the subject that deals with the different methods a system can use to adapt to the changes in the environment and continue to perform effectively. The idea behind this is that the system will need to anticipate a change in the environment and adapt to that accordingly.
Psychologists have been trying to understand the learning processes in human beings for quite some time. An example of one model used to understand these processes is the EPAM network. This network is used by human beings to store and retrieve one or two words when the other is provided. The concept of semantic networks and decision trees were later conceived in this field of learning. In recent times, the advances made in psychology have been influenced strongly by artificial intelligence. Another subject in psychology that is now being studied is called reinforcement learning which is a concept being used extensively in machine learning. This is covered in detail in later chapters.
A large chunk of machine learning is associated with artificial intelligence. Many studies on artificial intelligence are focused on the usage of different analogies to enhance the process of learning and also on how past events and experiences can be used to anticipate and adapt to future events. Recent studies focus on devising rules for systems using concepts of decision trees and inductive logic programming.
Evolutionary studies have concluded that animals not only learn to perform better in life but also try to learn to adapt to the change in their environments to enhance their performance. When it comes to machines, the concepts of evolving and learning can be considered synonymous. Therefore, models that have been used to explain and understand evolution can now be used to derive some machine learning techniques. The most prominent technique that has been developed using evolutionary models is the genetic algorithm.