The SIM neural module: self-organized learning of 2D invariant representations
Danny Roobaert, Marc M. Van Hulle · 1998
Developing spatial invariant 2D representations in a self-organizing way is realised by the Self-organizing Invariant representation Map (SIM) module which we introduce in this contribution. The model consists of a bi-layer structure of abstract neurons with adaptive forward connections and integrates weight-sharing with existing competitive learning rules. Adopting an unsupervised learning strategy, learning is fast and local in nature and implies that the model can be used as a modular building block at different levels of processing in larger vision systems. We discuss the behavior of the model on a toy problem and we demonstrate the integration of several SIM modules into larger application systems by reporting on a SIMbased handwritten numeral recognition system and a SIM-based natural object recognition system. Key words Invariant representation, unsupervised learning, neural networks, object recognition 1 Introduction Visual object recognition poses a number of challenges. One...