Novelty detection using products of simple experts—a potential architecture for embedded systems

The ‘Product of Experts’ architecture (concentrating on binary-stochastic elements) is described in the context of its suitability for implementation as mixed-mode hardware, with algorithmic modifications that render the training procedure hardware-implementable. Results show that the PoE is capable of modelling non-linear, multi-dimensional data drawn from both artificial and real sources. The capability of the PoE to perform on-line novelty detection is described and demonstrated on both artificial and real data.