29 March 2016, CLClab Seminar, Willemijn van Woerkom

The title of my research, and my talk, is "Modelling the Visual Number Sense Using a Deep Generative Neural Network". For my research project I have replicated a study done in 2012 by Stoianov and Zorzi: "Emergence of a 'visual number sense' in hierarchical generative models"(http://cognitrn.psych.indiana.edu/rgoldsto/courses/cogscilearning/). The authors trained two stacked Restricted Boltzmann Machines on reproducing simple binary images containing varying numbers of rectangular objects. After training, this model was able to sustain numerosity estimation by linear classifiers fed with its output. I have not been able to replicate all their results, and offer the likely causes for this.