Noise reduction in the nervous system. Focus on "Enhancement of ITD coding within the initial stages of the auditory pathway".

A frustrating reality of neuroscience is the tendency of neurons to respond in different ways to repeated presentations of identical stimuli. This variability might arise from several different sources, but to some extent, particularly at the level of sensory transduction, it is noise in the purest sense. In principle, the effects of noise can be mitigated by either averaging across time or across a population, or prior knowledge (Faisal et al. 2008). While these mechanisms are undoubtedly in use, there are few parts of the nervous system where noise reduction is the only thing happening. Instead, neurons presumably employ these noise reduction mechanisms in the course of performing some other computation. Where noise reduction as a computation by itself is found, it seems to occur primarily in the earliest stages of sensory processing: for example, in retinal ganglion cells (Balboa and Grzywacz 2000; Warrant 1999) or in the cochlear nucleus (Joris et al. 1994a,b) or following the electroreceptor primary afferents in electric fish (Carr et al. 1986). In this issue of Journal of Neurophysiology, Pecka and colleagues demonstrate the existence of noise-reduction-only neurons within the interaural time difference (ITD) pathway of the gerbil. ITD is a sensory cue for the position of a sound source, but because it is computed by detecting coincidences in spike timing between inputs arising from the two ears, it is a derived cue that is computed in the medial superior olive (MSO), three synapses removed from the cochlea. Qualitatively, there is very little difference in the tuning to ITD of neurons in the MSO from those in the dorsal nucleus of the lateral lemniscus (DNLL) (Kuwada et al. 2006) and later stages (Fitzpatrick et al. 1997). However, Pecka and colleagues show that compared with MSO, DNLL neurons have greater dynamic range in their response. In addition, the variability of DNLL neurons does not increase with mean firing rate; in fact, the ratio of the variance in spike counts to the mean spike count actually decreases with mean rate. By presenting their stimuli at intensities, they established that these differences predominately occur at higher intensities. Because the input to MSO neurons increases its firing with intensity, louder sounds lead to more spurious coincidences being reported and a degradation in ITD tuning. The net result is that single neurons in the DNLL encode greater information about the ITD of a sound than do single neurons in the MSO: the same sensory representation but reduced in noise. So in the gerbil, this derived sensory cue is followed by a stage of noise reduction reminiscent of the arrangement observed in primary sensory transduction. The question then is: to what extent does this reflect a general pattern in neural computation, as opposed to some unique quirk of the gerbil? While DNLL has not been studied in this light in other mammals, a similar explicit stage of noise reduction for ITD representation has been described in the barn owl (Christianson and Peña 2006; Fischer and Konishi 2008). The barn owl is extremely good at sound localization and unlike the gerbil—or indeed, any other described animal—has a specialized auditory system allowing it to use ITD cues at frequencies 10 kHz. That computation of the ITD should immediately be followed by improvement in the representation in both the gerbil and such a specialist is intriguing. This becomes especially true as the ITD pathways in mammals and avians are most likely to be analogues rather than homologues (Grothe 2000). If gerbils and owls have independently evolved similar neural structures for improving the representation of ITDs, it argues that this improvement of sensory representation arises from some fundamental principle of neural computation.