Rate coding versus temporal order coding: a theoretical approach
Introduction
For decades, neurophysiologists have characterised neural activity by the firing rate: PSTHs, tuning curves and even more recent stimulus-reconstruction methods all rely on stimulus-dependent changes in firing rate. Furthermore, the rate coding hypothesis has undoubtedly shaped the development of the ideas underlying artificial neural networks and PDP models.
However, recent studies that have looked at the speed with which sensory systems can process information (Thorpe et al., 1996) pose serious problems for traditional rate coding schemes. For instance, face selective neurones in primate inferotemporal cortex can respond only 80–100 ms after stimulus onset (Oram and Perrett, 1992). The fact that the first 5 ms of their responses is already selective suggests that the selectivity can be produced by essentially feed-forward propagation of information from the retina to IT via the LGN, V1, V2 and V4, a processing sequence involving roughly ten synaptic stages. On average, this leaves no more than 10 ms between two successive stages of synaptic activation, i.e. 10 ms for synaptic transmission, PSP conduction and integration, spike generation and spike conduction. We have argued that this period is too short to allow rates to be determined accurately, because few neurones will fire more than one spike in this time (Thorpe and Imbert, 1989).
Section snippets
Rate coding of analog values
In this section, we will examine the efficiency of rate coding using a very simple Poisson process for generating spikes. It is clear this is only a very approximate model for the firing of real neurones. However, we believe that the general points that can be drawn will apply to any scheme relying solely on counting the number of spikes.
From the efferent neurone’s point of view, the less time there is to evaluate afferent spike rate, the more inaccurate this evaluation. If we assume that spike
Comparisons between two analog values
Neural processing is often aimed at detecting differences in activation rather than absolute values: contrast rather than pure luminance, edges rather than areas, etc. Would a rate coding strategy be more efficient for such a purpose?
Let us take two populations A and B, each composed of N neurones, which send spikes to a second-level neurone D that has to react on the basis of the difference of activity between the two afferent sub-populations. For instance, the efferent neurone D could test
Asynchrony: another way to code
In this section, we will consider some of the alternatives to rate coding. Typically, these involve some form of temporal coding, i.e. a code in which the timing of spikes plays a crucial role. With very short time scales (10 ms or less), time coding and rate coding tend to become confounded since ultimately, any form of temporal code can be described in terms of very rapid changes in firing rate (Rieke et al., 1997). However, the situation becomes much more interesting if we consider the
Rank order coding
One simple way of using asynchrony is to use the order in which the neurones spike as a code. In this case, the exact latency at which a neurone fires is not critical—only the rank order of each neurone is important (Thorpe and Gautrais, 1997; Thorpe and Gautrais, 1998). Such a scheme offers a number of advantages.
Firstly, a code based on the order will be more robust to noisy temporal jitter of each spike than a pure temporal code that must rely on temporal precision, especially when decoded
Conclusions
The idea that neurones transmit information in the form of a rate code is extremely entrenched. There have been numerous other suggestions over the years (Perkel and Bullock, 1968), but they have done little to overturn the overwhelming popularity of the rate coding hypothesis. Even today, with more and more researchers interested in the possibility that temporal synchrony might play an important role in neural computation (Abeles, 1991 Singer and Gray, 1995), most people still consider that
References (21)
Variation in the response latency of cat retinal ganglion cells
Vision Res.
(1973)- et al.
Visual latency of X- and Y-cells in the dorsal lateral geniculate nucleus of the cat
Vision Res.
(1986) - et al.
Face processing using one spike per neurone
Biosystems
(1998) - et al.
Synaptic depression and cortical gain control
Science
(1997) - Abeles, M., 1991. Corticonics. Neural circuits of the cerebral cortex. Cambridge University Press,...
- et al.
Responses of neurons in primary and inferior temporal visual cortices to natural scenes
Proc. R. Soc. London
(1997) - et al.
Asynchrony in visual analysis: Using the luminance-to-response-latency relationship to improve segmentation
J. Opt. Soc. Am. A
(1994) - et al.
Dynamics of orientation coding in area V1 of the awake primate
Visual Neurosci.
(1993) - et al.
Adjacent visual cortical complex cells share about 20% of their stimulus-related information
Cereb Cortex
(1996) - Nowak, L.G., Bullier, J., 1997. The timing of information transfer in the visual system. In: J. Kaas, K. Rocklund, A....