Wed 4 Oct 1995, 12:00, WeH 7220

	       Logistic Regression for Classification:
	       Can it be an alternative to neural nets?

			       Kan Deng

A binary-output classification question can be transformed into
seeking proper parameters of the logistic function.  In statistics,
this is called GLM.  This method can be easily generalized to the
multiple-output case.  However, the GLM approach does not separate
complex classes well.  In this talk, we explore and evaluate two
extensions: Locally Weighted Logistic Regression and Composite Feature
Logistic Regression.

Compared with neural nets and other machine learning methods, these
new techniques are explicit in expressing the relationship of the
input attributes and outputs; they do not require a separate training
phase; and they offer solid statistical confidence measurements for
their predictions.  With all these properties, these new techniques
are especially promising for online autonomous classification system.