On Grandmother Cells

Grandmother cells were proposed as single neurons which encoded for a single concept - in this case one's Grandmother. The concept was used as a straw-man to criticise sensory hierarchy-based brain organisation theories, but the ideas underlying it are becoming increasingly accepted (not the single neuron per concept I hasten to add). In a paper by Charles Gross (2002 - full ref below), a history of the term is given, with origins and influences.

These cells were originally proposed at the opposite end of the spectrum to ensemble or population coding - where it is the pattern of activity across a group of neurons which codes for a sensory percept. Although the term was coined by Jerry Lettvin in the late 1960's, after which its use quickly proliferated, it was actually proposed as a scientific theory a number of years earlier by the Polish neurophysiologist Jerzy Konorski.

Proposed in his work "Integrative Activity of the Brain" (1967), Konorski predicted the existence of individual neurons sensitive to sensory stimuli such as faces, hands, emotional expressions, etc - and named them "Gnostic" neurons. These were proposed to be located in specific areas of the cortex (in "gnostic fields"), such as the ventral temporal cortex (for the face field), and the posterior parietal cortex (for space fields). These predictions have proven reasonably similar to current proposals for the extra-striate visual cortex in monkeys.

Naturally however, Konorski's work was influenced by that of others. Firstly, in the early 1960's, Hubel and Wiesel demonstrated the hierarchical processing of sensory information in the geniculo-striate system: from simple receptive fields up to the ability to selectively generalise across the retina. Secondly was research on what was then known as the Association Cortex by Pribram and Mishkin: lesions of the inferior temporal cortex produced specific visual cognition impairments in monkeys.

These two bodies of evidence, along with his own familiarity with various agnosias which follow human cortical lesions, led to Konorski's proposal of gnostic cells as a means of accounting for these cognitive impairment effects. Despite the publication of these ideas, and the subsequent coining of the term 'grandmother cells', for at least a decade afterwards, gnostic cells were only taken up in the learning literature, not the perception literature. The term has now, however, had greater use in general textbooks and the pattern recognition literature.

Two features of the gnostic cells have long histories in neuroscience research. Firstly, they are examples of labeled line coding. Labeled line coding refers to a neuron property that allows it to code a particular stimulus property, such as line orientation in the visual field. Secondly, gnostic cells were held to be at the top of a 'hierarchy of increasing convergence'. This concept of convergence hierarchies had, for example, been proposed by William James (the pontifical cell), C.S. Sherrington (in "Man on His Nature", 1940), and Barlow (with the slightly modified concept of cardinal cells, 1972).

In conclusion, the paper notes that the idea of convergence of neural input onto one cell seems to have arisen independently a number of times - and that contemporary human brain imaging have revealed cortical regions (e.g. inferior temporal cortex) which resemble the gnostic fields proposed by Konorski.

As an example of more recent research efforts in exploring this converging hierarchy proposal, Quiroga et al ("Sparse but not 'grandmother-cell' coding in the medial temporal lobe", TICS, 12(3), 2008) - which also involved the somewhat infamous experiments which identified the 'Jennifer Aniston cells' - identified very sparse coding of visual percepts in the medial temporal lobe. In this paper though, a number of arguments were presented for why these cannot be considered to be grandmother cells - a view which I think may be widespread: sparse encoding but not convergence onto a single cell.

Gross, C. (2002). Genealogy of the "Grandmother Cell". The Neuroscientist, 8(5), 512-518.