Tactile feature extraction and classification with connectionist models

Abstract

Interim results of a study on pattern recognition of robotic tactile impressions using connectionist models are described. The training data consists of gray-scale force gradient profiles that accurately reflect the tactile domain; the focus is on extracting these features with artificial neural systems (ANSs). A description is given of an architecture in which a two-layer back-error-propagation network performs feature extraction of gray-scale gradients, and a second BEP network classifies the surface profiles. Imposition of constraints on the training set is critical to ensure that meaningful features are selected. In domains where information content of the input vectors are dense and very similar, receptive field neurons encode useful data across unit activations, while fully connected schemes shroud information among the link weights