Monday, 22 March, 2004:

<CENTER><B> Do neurons need a large dynamic range?</B>

<P>
Matthias Bethge <BR>
Redwood Neuroscience Institute
</CENTER>

Several aspects of distributed neuronal representations (population 
codes) can be optimized with respect to coding efficiency. While the 
efficient coding hypothesis (Attneave, Barlow) emphasizes the 
adaptation of the population code to the sensory signal statistics, 
there are other factors affecting the coding efficiency  of neuronal 
representations as well. Coarse coding for instance has been proposed 
as an efficient coding strategy for populations of neurons (Hinton, 
1981). While the original argument for large tuning widths has been 
built upon a binary neuron model, population coding is commonly thought 
of rather in terms of analog rate coding, mostly due to the fact that 
the large majority of measured tuning functions look smooth and 
bell-shaped. Later results on optimal tuning widths for smooth 
bell-shaped tuning functions differed from those of the binary case. 
Using a more general model which allows to vary the dynamic range of a 
tuning function independently from the tuning width, the divergent 
conclusions in the literature can be explained in a unique way. The 
optimization of the tuning width can be understood to a large extent in 
terms of the effect of the individual firing rate distributions on 
coding efficiency, which is strongly regularized if the mean firing 
rate is assumed to be given. The dynamic range, however, seems to be 
less predetermined. On the basis of analytical and numerical studies I 
show that a minimal dynamic range is optimal in terms of coding 
efficiency.  In conclusion, the efficient coding principle argues 
against the idea of analog rate coding (smooth tuning).