The Maximal Causes of Binary Data
Jacquelyn A. Shelton · 2010
Neural activity encodes multiple-cause stimuli with discrete events.Neurons either spike or remain inactive. Many modeling approachestherefore rely on binary units for encoding. Prominent examples are,for instance, restricted Boltzmann machines [5] and, more recently,deep belief networks [6]. In this work we study a probabilisticgenerative model with binary units. We investigate the componentextraction capabilities of a model with hidden and observed layerboth encoding binary data through Bernoulli distributions. In thissetting basis functions can not be combined using summation as insparse coding models [3] but require non-linear combination rules.