In recent years, several studies have examined decision making on the basis of ambiguous visual stimuli. Neural recordings during a task of motion discrimination have shown that activity in the lateral intraparietal area (LIP) increases proportionally to the strength of visual evidence. This accumulation, however, is observed in the average population activity but not in individual neurons, which display a wide heterogeneity of response patterns (also termed response "motifs") [1]. It remains unclear what functional role may be linked with this neural heterogeneity. Here we propose that heterogeneity promotes the accurate discrimination of input through a transformation termed a high-dimensional projection. This effect is shown in a simple linear rate model with balanced excitation/inhibition and orientation-selective units. In the model, a balance of excitation and inhibition allows neuronal activity to produce complex, yet globally stable activity. While the maximum amplitude of responses to stimuli varies between units, the average variance across individual units is several orders of magnitude lower than the variance of average population activity. This behavior of the model allows population activity to act as a line attractor despite fluctuations in single-unit activity. The pattern of activity produced in the model forms intricate response motifs that differ based on input statistics. We suggest that these response motifs can be used to perform stimulus classification based on the target decision of individual trials in the dot-motion task. We analyzed the activity of the model using a Fisher linear discriminant analysis, and show that units with uncorrelated activity project the activity of the model to a high-dimensional space where a linear manifold can classify stimuli accurately. This high-dimensional projection has been studied in other contexts including object recognition [2], and can be read out linearly using a simple Hebbian rule. We show that neural correlations induce a fundamental limit on the readout capacity of the model by inducing collinearity amongst stimulus responses. As a result, the model performs below theoretical bounds for classification accuracy. We employed the model to capture neuronal responses in recordings of LIP in monkeys performing a task of visual motion discrimination. The model generated five novel predictions that were confirmed in analyses of electrophysiological data: (1) a linear relation between classification error and stimulus coherence; (2) a gradual decrease in classification error as more units are incorporated in the high-dimension projection; (3) an above-chance classification based on out-response field neurons alone; (4) higher classification accuracy using in-response field neurons than out response-field activity; and (5) chance-level discrimination with the removal of linear accumulation. In sum, results provide an alternative to the common notion that accurate neural representations of input rely on a simple accumulation of population activity to a set threshold, and instead suggest that heterogeneous activity plays a key role in generating a high-dimensional projection where linear discrimination attains a high accuracy.

[1] M.L. Meister, J.A. Hennig, and A.C. Huk, Signal multiplexing and single-neuron computations in lateral intraparietal area during decision-making. J Neurosci, 2013. 33(6): p. 2254-67.
[2] M. Rigotti, et al., The importance of mixed selectivity in complex cognitive tasks. Nature, 2013. 497(7451): p. 585-90.