Mathematical Neural Network (MaNN) Models Part III: ART and ARTMAP in OMNI_METRICS
K. Ramakrishna, Veluri Anantha Ramam, R. S. Suriavel Rao · 2014
(Dedicated to Dr U Murali Krishna, former Special officer, A U P G center, Nuzvid, former Professor of [Analytical, engineering] department, Andhra University, on his sahasra chandra darsanam (thousand lunar months of life on the lap of mother earth). _____________________________________________________________________________ ABSTRACT Adaptive resonance theory (ART) proposed by Grossberg in 1976, is a self-organizing (SO) unsupervised learning approach. It balances stability vs. plasticity dilemma in learning new traits without forgetting the old ones. Another popular SO mapping (SOM) of Kohonen introduced in 1990s subtly differs from ART and based on neighborhood influence. ART1-NN is the start of a new era of unsupervised-data-driven models using resonance, ordinary differential equations (ODEs) and backward connections. The main functioning of ART1 is in feature and category layers which are connected both ways. It accepts only binary input and makes use of winner-takesall (WTA) and vigilance approaches. ART2 and fuzzy-ART are modifications of ART1 to handle analogue and floating point input values. ART3 performs a parallel search and incorporates a term similar to chemical transmitters playing a key role in biological system. The NNs reported over two decades in this category include coupled-, probabilistic-, projection-, performance-guided-, lateral-priming-, efficient-, fusion-ARTs. The philosophy of Grey relational analysis inspired from human brain is used in Grey-ART. Multiple channel data is analysed with multi-ART. Under hybrid category, SOM is used in fully-organized-SOM-ART and RBF in RBF-ART.