08 January 2012. Nature Neuroscience (2012). doi:10.1038/nn.2996 ---The ability to approximate (estimate), rather than calculate, can be critical in complex situations. It might also explain why invention of number Zero was so controversial. We have a hard time "seeing", i.e. creating an internal neural network representation, a non-existing pattern.
Here we show that visual numerosity emerges as a statistical property of images through unsupervised learning. We used deep networks, multilayer neural networks that contain top-down connections and learn to generate sensory data rather than to classify it8, 9. Stochastic hierarchical generative models are appealing because they develop increasingly more complex distributed nonlinear representations of the sensory input across layers9. These features make deep networks particularly attractive for the purpose of neuro-cognitive modeling.
tags: pattern, approximation, science, network, biology