Special
Three questions to Professor Ryszard S. Michalski, founder of Machine Learning Science and author of 16 books and over 350 publications.

His biography and work on
http://www.mli.gmu.edu/michalski/michalski.html

1. ELKM, Europeen Community Special Interest Group (SIG) was created to discuss the topics as education and learning for the Knowledge Management. According to our discussions all human knowledge processors have to learn in real time from interactions with the external world. A machine such as computer, cellular phone, PDA have a role to play in the process of learning. Since over 30 years now you have been involved in Machine Learning. Could you explain how computer can learn and what are the best methods to do it well ?

RSM :
Computer learning is limited by the type of knowledge representation used for implementing learning processes, and the type of operators it can apply to learn new knowledge using this knowledge representation. During the last 30 years, there has been a great amount of effort expended in the area of machine learning, and a lot of methods have been developed, but a relatively little progress has been achieved as to the types and the power of knowledge representations that the learning programs are able to use, and the type of "knowledge generation operators" or the complexity of inference that they can perform in order to learn, or to update previously learned knowledge.

For example, one of the simplest learning methods, decision-tree learning, has been originally developed in the early 60-ties, but still remains one of the most popular today. Compare this with using, at present, an over 40 year computer! The creation of a decision tree is an application of a relatively simple operator of attribute section at each consecutive node of the decision tree under development. The method does not invent or involve higher-level concepts that not present in the input data, does not create new knowledge on the basis of the previously learned knowledge, and does not represent concepts in a flexible and context-dependent way, that is, lacks the capabilities typical of human learning. Typically, humans give a computer program a dataset, usually in the form of attribute-value vectors, and the program builds a tree structure that can classify the input vectors, and future vectors into the previously predefined categories. Despite its simplicity, this method has many practical applications. But it is a very simple form of learning.

There has been a tremendous amount research done on the development of various methods of machine learning, to wit, rule learning, support-vector machines, Bayesian net learning, inductive logic programming, neural nets, nearest neighbours, genetic-algorithms-based classifiers, etc., which found a lot of useful applications. And there will be more progress and more useful applications in the future. All these methods, however, with a (partial) exception of inductive logic programming and some special cases, basically perform a simple "open-loop learning" that take input data, and build a structure that generalizes (better or worse) the input data. These methods can not learn or develop new complex knowledge by utilizing the previously learned knowledge (that is, are not able to perform a "closed-loop" learning).

In short, despite much effort in the last 30 years, there has not been much progress in the conceptual scope of capabilities of learning systems that would bring them truly closer to human learning, in particular, to closed-loop learning. This does not mean, however, that there has not been a significant progress and many practically useful results. The problem is really very hard, and much effort is required to make further progress. But its fruits can be very sweet.

2. Do you believe that it is possible to have real and efficient learning machine, a human-like machine ? What are the conditions ?

RSM:

Human learning is a process of creating or updating knowledge using input sensory information and past knowledge---it is arguably the most intellectually complex process in our universe, as we know it. Developing and implementing a full computational model of human learning in a machine or a robot, or perhaps even a more powerful model of learning than that, will change dramatically the human society. Imagine an "intelligent robot" that can in an hour (or a day, or a week) read, learn, and then reason with all the knowledge contained in the Encyclopaedia Britannica, or read and comprehend a thousand books during such a time, remember their content for its lifetime, and then flexibly use it for making decisions or solving problems. How humans will be able to co-exist and interact with such an intellectual power?

Don’t worry, developing such a robot may take a very long time, perhaps 50-250 years of a peaceful human existence, but science and technology are striving in this direction. Such a system will not, certainly, be built in our lifetime. But in the future--nobody knows. One cannot really tell either way at this time. The humanity will in the meantime face much more down-to-the-earth problems, which it needs to solve in order to survive.

3. How do you imagine the capacity of a machine in 2020 ?

RSM :

It is actually quite amazing to see such an immense progress in the last thirty years in the development of computer hardware, information technology, and the internet, as compared to a relatively modest progress in endowing computers with capabilities to learn and reason. This simply confirms what some of us suspected for some time, that developing powerful learning machines is one of the hardest, if not the hardest, scientific challenges facing humanity. At this time, learning systems are far away from having a capability to incrementally learn (or improve) complex knowledge from different sources (e.g., by reading books or websites), and then use it in their reasoning.

Judging by the progress in machine learning in the last 30 years, I do not expect a very great progress in developing significantly greater capabilities before 2020, that is, in the next 15 years. A major breakthrough will appear when reusable knowledge bases will be widely available to researchers, and they will learn how to utilize them effectively and efficiently in machine learning/computational learning or data/knowledge mining systems. Research and development of such systems is and will remain a fascinating and practically very important research goal for a very long time.

Thank you Ryszard