Title: Grandmother cells and the storage capacity of the human brain

Abstract: Quian Quiroga et al. [Nature 435, 1102 (2005)] have recently discovered neurons that appear to have the characteristics of grandmother (GM) cells. Here we quantitatively assess the compatibility of their data with the GM-cell hypothesis. We show that, contrary to the general impression, a GM-cell representation can be information-theoretically efficient, but that it must be accompanied by cells giving a distributed coding of the input. We present a general method to deduce the sparsity distribution of the whole neuronal population from a sample, and use it to show there are two populations of cells: a distributed-code population of less than about 5% of the cells, and a much more sparsely responding population of putative GM cells. With an allowance for the number of undetected silent cells, we find that the putative GM cells can code for 10^5 or more categories, sufficient for them to be classic GM cells, or to be GM-like cells coding for memories. We quantify the strong biases against detection of GM cells, and show consistency of our results with previous measurements that find only distributed coding. We discuss the consequences for the architecture of neural systems and synaptic connectivity, and for the statistics of neural firing.