Finally, it is unclear whether the effect of training on correlated noise is specific to tasks for which area MSTd is thought to provide critical input. If we had trained Selleck BAY 73-4506 animals to perform a task that was irrelevant to self-motion perception, such as a somatosensory or auditory discrimination task, we presumably would not expect to see changes in correlated noise in MSTd. However, this possibility remains to be tested. Despite a robust effect of training on the average noise correlation in MSTd, our simulations show that an optimal, unbiased decoding of all neurons does not predict

a substantial change in performance due to learning. Indeed, theorists have shown that correlated noise may or may not harm population PI3K inhibitor coding (Abbott and Dayan, 1999, Averbeck et al., 2006 and Wilke and Eurich, 2002). In general, positively correlated noise between neurons with similar tuning (or more generally, any situation in which both neurons fire more strongly under one stimulus/task condition than another) harms the signal to noise ratio of the population code because it cannot be removed by pooling across neurons (Bair et al., 2001, Shadlen

et al., 1996 and Zohary et al., 1994b). Reducing shared noise among neurons in such cases is thus expected to improve population sensitivity. Indeed, the effect of attention on the fidelity of population codes appears to

follow this logic (Cohen and Maunsell, 2009). In a typical spatial attention task, most neurons with receptive fields at the attended location will increase their response. Because attention has a consistent polarity of effect on the responses of nearby neurons, stronger attention will tend to increase the responses of both neurons in a pair. Hence, most pairs of nearby neurons will have positive signal correlations with respect to the effect of attention. As a result, a reduction in correlated noise due to attention can improve the signal-to-noise ratio of the population code. However, in other contexts whatever for which signals are decoded from populations that include neurons with dissimilar tuning properties, increasing correlated noise can improve the signal-to-noise ratio of a population code (Figure 7A), as differences in tuning effectively cancel more of the noise in a population response (Abbott and Dayan, 1999, Averbeck et al., 2006, Poort and Roelfsema, 2009 and Wilke and Eurich, 2002). Reducing correlated noise in the latter case can harm the coding efficiency of the population. In our heading discrimination task, it is likely that responses are decoded from neurons with a broad range of heading preferences (Gu et al., 2008b and Gu et al.