Classification of incomplete data using the fuzzy ARTMAP neural network

A What-and-Where fusion neural network for recognition and tracking of multiple radar emitters

“…Training on such data is referred to as "semisupervised learning" (Demiriz et al, 1999) or "partially supervised clustering" (Bensaid, 1996;Pedrycz, 1985). To assess the effect on performance of training fuzzy ARTMAP using data with missing class labels, the network was trained in two phases (Granger et al, 2000). During the first phase, involving supervised learning, the network was trained as usual until convergence with a fixed amount of labeled training data from each radar type.…”

“…During the second phase, involving unsupervised learning, the network was presented with a varying percentage of unlabeled data from each radar type until the weights W converged once again. Using fuzzy ARTMAP without the class prediction (i.e., without Step 4 of the fuzzy ARTMAP algorithm), plus other modifications discussed by Granger et al (2000), the network associated each unlabeled training pattern with one of the already-existing Fz category nodes and adjusted the corresponding prototype vectors through slow learning (0 < P < 1). In all simulations with the radar data, the classification rates observed were never greater than those achieved by simply discarding all unlabeled data (Granger et al, 2000).…”

“…Using fuzzy ARTMAP without the class prediction (i.e., without Step 4 of the fuzzy ARTMAP algorithm), plus other modifications discussed by Granger et al (2000), the network associated each unlabeled training pattern with one of the already-existing Fz category nodes and adjusted the corresponding prototype vectors through slow learning (0 < P < 1). In all simulations with the radar data, the classification rates observed were never greater than those achieved by simply discarding all unlabeled data (Granger et al, 2000). Such an approach is most effective to the extent that clusters of data from different emitters are separable and well clustered, which is not necessarily the case for radar ESM data.…”

“…A strategy to address missing components in the input patterns consists in using indicator vectors (Ghahramani and Jordan, 1994;Granger et al, 2000;Little and Rubin, 1987). An indicator vector 8 = (8t,8z, ... ,8M) informs the fuzzy ARTMAP network about the presence or absence of each component in an input pattern: 8; = 1 if component i is present, and 8; = 0 if component i is missing, with 8;+ 111 = 8; fori= 1, ... ,n1.…”

“…An indicator vector 8 = (8t,8z, ... ,8M) informs the fuzzy ARTMAP network about the presence or absence of each component in an input pattern: 8; = 1 if component i is present, and 8; = 0 if component i is missing, with 8;+ 111 = 8; fori= 1, ... ,n1. Unlike strategies that involve replacement by "0" or by "1" (Granger et al, 2000), the indicator vector strategy modifies the prototype vectors as well Learning:…”