Abstract
The aim of the paper is to combine the generalization power of classical density estimation techniques with the efficiency of VLSI-friendly, constructive algorithms. This is important for VLSI-implementations of classification networks in the case of noisy data to avoid the well-known “overfitting” effects. The method consists of three steps 1) estimation of the probability densities for each class, 2) discretization of the input space and 3) describing the resulting classification regions using only easy implementable boolean AND and OR gates and comparisons. The “noisy spirals” classification problem, a noisy variant of the “two spirals” benchmark, is used for demonstration.
Original language | English |
---|---|
Title of host publication | Proceedings of the Third Annual SNN Symposium on Neural Networks |
DOIs | |
Publication status | Published - Sep 14 1995 |
Event | SNN'95 - Nijmegen, The Netherlands Duration: Sep 14 1995 → … |
Conference
Conference | SNN'95 |
---|---|
Period | 9/14/95 → … |