Paper Type |
Opinion |
Title |
From Natural Classification to Highly Intelligent Machine Learning |
Author |
Chidchanok Lursinsap |
Email |
- |
Abstract: 1. NATURAL SURVIVAL
The process of classification has existed in the nature since the beginning of life. At the beginning, every living creature adapted its own classifying mechanism to select the edible food. Since then this primitive classifying mechanism has gradually developed itself through the biological evolution to cope with more complicated situations for survival such as decision making, designing tools, seeking knowledge, and enlightenment. In human society, any one having high correctness and high versatile ability to perform the classification is considered as a highly intelligent person. Nowadays, humans cannot survive without any help from machines. Intelligent machines are new species emerging in this era of technological evolution. The ultimate goal of developing these intelligent machines is to release humans from tedious jobs so that humans can concentrate on the advanced scientific discoveries and gain happiness. The interesting problems is how to capture the natural classification in forms of mathematical models and implement the models in these machines to make them intelligent.
2. LEARNING VERSUS CLASSIFYING Suppose an orange and a banana are placed on a table and a child is told pick a round shaped object laying on the table. What is the first procedure the child must do in order to make the decision ? The firrst step is to scrutinize the objects on the table and extract the features relevant to the shapes of the objects. Then, these extracted features are classified into either round class or not round class. This testing is easy for a child since the shapes of the orange and the banana are obviously distinctive. But if a water melon is also placed on the table then the decision is confused. The melon has an elliptic shape which is not exactly round or long. To make the correct decision, the child must learn all possible features of the objects considered as round objects prior to the classification. Thus, learning and classifying are two inseparable processes. This observation is adopted in machine learning. In machine learning, there are three types of learning, i.e. supervised, unsupervised, and reinforcement. In supervised learning, a machine is trained and forced to produce the desired output or target corresponding to its training input pattern. For example, if the image of an orange or a durian is presented to a trained machine, then the machine should produce the output of value 1 denoting that the presented image has a round shape. In unsupervised learning, no target is desired. The machine groups the set of patterns into classes by using some similarity measure. An example of this type of learning is shown in Figure 1. The objects are classified into two classes according to theirs similarities. Here, the similarity measure is based on Euclidean distance among objects. In reinforcement learning, the machine is forced to produce the correct output by giving only rewards or penalties based on the generated output. This is similar to animal training where the direct communication between human and the animal cannot be established. 3. SOME RESEARCH DIRECTIONS IN INTELLIGENT MACHINE LEARNING The group of features of an object can be viewed as a vector in a high dimensional space. The dimension is equal to the number of distinct features. Typically, the vector is called feature vector. Hence, the problem of learning is transformed to the problem of partitioning the set of feature vectors into classes. The supervised learning and unsupervised learning are two most studied problems. In case of supervised learning, the class of each feature vector is known prior to the partitioning process. The research competition in this area is finding new nonlinear separating functions and a set of kernel functions for mapping the feature vectors into a higher dimension that produce minimum number of misclassified feature vectors. Generally, any non-parametric separating functions can be realized by a neural network. In case of unsupervised learning, since the classification is mainly based on the similarity measure, the competition will be concentrated on a new similarity measure in forms of non-Euclidean distance. Besides the pure learning research, the applications of machine learning must be emphasized as well. The problems in the areas of bioinformatics, pattern recognition, natural phenomena simulation and modeling, animation, data mining, drug design, and decision support system can be transformed to the problem of learning and classification. |
|
Start & End Page |
247 - 248 |
Received Date |
|
Revised Date |
|
Accepted Date |
|
Full Text |
Download |
Keyword |
|
Volume |
Vol.33 No.3 (SEPTEMBER 2006) |
DOI |
|
Citation |
Lursinsap C., From Natural Classification to Highly Intelligent Machine Learning, Chiang Mai J. Sci., 2006; 33(3): 247-248. |
SDGs |
|
View:708 Download:159 |