Differences between colour codes

It is known that many cells throughout the visual system are sensitive to coloured input. From the retina through to the LGN and up to cortical regions, V1 and V4, many neurons show selectivity to specific colours. However, exactly what these cells are coding for is still uncertain. There are 3 main possibilities.

1) Colour biases
A neuron may show colour selectivity as a secondary characteristic, and this this may be functionally irrelevant. For example, consider the case of orientation selective cells in V1, or motion selective cells in V5, whose primary job is to code for these properties. Just because the cells shows some colour sensitivity it does not mean that this information is used in a useful way by later stages. These biases may result form 'random cone clustering' (Conway et al., 2008). Consistent with this proposal is the fact that these colour biases change when the luminance of the input changes. This should not happen with true hue selectivity.

2) Wavelength selectivity.
Although we perceive a unified colour at any one place/time, the input is actually decomposed, by the 3 cone type photoreceptors responding to different wavelengths. Cells near the bottom of the visual system (the retina and LGN) appear to code for colour in this way. For example, seeing yellow is generated by the correct ratio of 'red' and 'green' wavelength receptor activation, although there is no hint of red or green in the banana! This lead onto the third possibility

3) Perceptual colour / hue selectivity.
It is logical to expect that at some stage there exist neurons whose activity correlates with perceived colour. At this stage a 'yellow' neurons would respond to our banana.

A number of possibilities have been raised here. Evidence suggests there is a dissociation between wavelength and perceptual colour codes, with the former associated with early visual areas (retina, LGN, and V1) and the later with V4 (refs).

Brouwer & Heeger (2009) have recently probed colour codings in different areas using the multivariate pattern analysis approach (MVPA). This is an exciting new method for teasing apart spatial overlapping neural representations using conventional brain scanning measurements. Essentially the computer is given the results of a brain scan and asked to do its best to dissociate between different classes of activity. The computer is able to use some pretty fancy statistical techniques so does a far better job than a old fashioned human eyeballing approach could hope for. They show that the classifier is able to correctly predict the stimulus colour, based on activity patterns in v1,v2,v3,v4,vO1,LO1. However, only in V4 and VO1 is a gradual change in perceptual hue mirrored by a gradual change in neural activity patterns. Thus these areas are most likely strongly involved in perceptual coding of hue.

Tying these results into the 3 possible signal types described above it would be very interesting to if the classifier is, in some cases relying of the first type of signal. That is colour bias signals, which in this cases would be an artefact. Brower & Heeger (2009) report best performance using the activity patterns in V1, but could this be due to a relatively large proportion of colour biased cells - that is, a colour signal that isn;t useful for perception. To determine the perceptual relevance of colour related activity in v1, the Brower & Heeger (2009) experiment could be rerun with each specific hue being presented at a number of different luminance levels. This may greatly reduce colour related activity in v1 because hue biases in v1 are often altered or abolished when stimuli are raised or lowered in luminance (Solomon & Lennie, 2007).

refs coming soon...