Facial Features Classification Using the Temporal Correlation Matrix Memory (TCMML)

Abstract

This paper examines the motivation and concepts of dynamic encoders (for binary neural networks) introduced by Shah in the author's RASC 2004 and 2006 conference papers [1,2]. Further to this, the paper extends the claims made by Shah et al. in their IJCNN2007 conference paper [3] about dynamic encoders and offers a different understanding to using dynamic encoders. In addition the paper also derives the Improved Correlation Matrix Memory (CMML) (first introduced by Shah et al.) via practical considerations (as opposed to a theoretical concept), supplies a theorem that provides the missing justification over the use of the improved adjective, before finally enhancing the CMML into the `Improved' Temporal Correlation Matrix Memory (TCMML) together with a brief discussion on an application for recognising facial features.