From A conversation about the brain
Jump to: navigation, search

Although it is not essential for the discussion to relate the ideas raised here to particular neural structures, it may help. The cerebellum is a vast store of synaptic weights which are set up in such a way that they can be readily compared with cortical inputs. Throughout evolution, there has been a highly preserved ratio between the surface area of the cerebral and cerebellar cortices [1]. The granule cells in the cerebellum -- as numerous as all the neurons in the cerebral cortex -- allow the cortical input to be compared with a much wider range of the cerebellar stored weights than would otherwise be the case. The cerebellum has, since the 1960s, been known as a place where a very high dimensional input vector ($\vec{r}$, see Notation - if Maths notation does not display correctly try refreshing the page) can be compared with a vast number of stored vectors in the same space ($\mathbf{W}$, see Notation). If the input is $10^{10}$ dimensional and the stored vectors are too, then the outputs must be encoded by many Purkinje cells (one Purkinje cell has only $10^{5}$ synapses).

The cortex needs to provide a motivational component ($\vec{t}$, see Notation) to the input vector, e.g. provided by the pre-frontal cortex and limbic system. The dorsal and ventral visual streams provide complementary elements in the sensory component of the input vector, carrying information about the movement of the observer and the constant visual aspects of the scene respectively. The hippocampus, an autoassociation network at the end of the ventral stream, aids the stability of the latter component.


  1. Sultan, F. (2002). Brain evolution (Communication arising): Analysis of mammalian brain architecture. Nature, 415(6868), 133-134.