Numerosity perception refers to the ability to estimate the number of items in a visual scene. This thesis explores numerosity perception in ‘Deep Belief Networks’ (DBNs), hierarchical generative models that learn the underlying statistical representation of the sensory input in an unsupervised fashion. To simulate behavioral tasks, the internal representations were read-out by a supervised linear classification layer. Networks’ performance was assessed using three numerosity judgement tasks, namely pairwise numerosity comparison, fixed-reference comparison, and numerosity estimation task. Furthermore, the influences of training dataset size and type of classifier used for read-out were systematically examined. Results showed that DBNs performed accurately across tasks and conditions, capturing the key behavioral patterns observed in human empirical studies.