Figure 1
From: Molecular function recognition by supervised projection pursuit machine learning
Schematic of SPLOC as a recurrent neural network and data flow. For p variables there are p perceptrons, labeled from 1 to p, comprising the input layer that receives \(N_F\) functional and \(N_N\) nonfunctional data packets of n samples. Each perceptron maps to a mode, and has access to all data packets organized in the form of two types of data packet cubes. Each perceptron recurrently interacts with all other perceptrons through competitive learning. The basis set rotates as the neural network evolves to maximize efficacy. Upon convergence, all perceptrons comprise the output layer for the specification of an orthonormal complete basis set. A rectifying function is assigned to each perceptron, defining a viewpoint for controlling sensitivity and selectivity in feature extraction. For a given viewpoint, the final basis set defines perception when the neural network achieves maximum efficacy. Unlabeled data packets are subsequently classified within the context of training data, having multivariate discriminant and conserved features that are readily interpretable.
