Incorporating Fuzzy Knowledge into Neural Networks


A priori Knowledge and Fuzzy Rules


Fuzzy logic is an efficient tool for representing fuzzy rules, such as heuristics in IF-THEN structures [1].


Data and Knowledge for Neural Network Learning

In many real-world applications of neural networks, it is not trivial to collect sufficient training data. On the other hand, there is usually domain knowledge, heuristics and expert knowledge available. Unfortyunately, conventional learning algorithms are not able to take advantage of the knowledge that is not in the form of data pairs. Therefore, it is very desirable to develop a method that is able to use data as well as knowledge .


Knowledge Incorporation through Regularization and Multi-task Learning

A priori knowledge can be ultilized in neural network learning in the following to ways:

  • Fuzzy knowledge as regularization. By representing knowledge with fuzzy rules, it is possible to regularize data-driven learning to speed up convergence and to improve accuracy.
  • Fuzzy knowledge as a related task. Multi-task learning is very attractive [2], however, it is not easy to define a related task. Knowledge represented with fuzzy logic can be treated as a related task.


  • A method that is able to incorporate knowledge with the help of fuzzy rules has been proposed in [3], which has shown to be effective in improving learning performance.

    References

    [1] L. Zadeh. Knowledge representation in fuzzy logic. IEEE Transactions on Knowledge and data Engineering, 1(1):89-100, 1989
    [2] S. Thrun and L. Pratt. Learning to Learn. Kluwer Academic Publishers. 1997
    [3] Y. Jin and B. Sendhoff. Knowledge incorporation into neural networks from fuzzy rules, Neural Processing Letters, 10(3):231-242, 1999


    For discussions, please contact Yaochu Jin.