In the middle temporal (MT) area of primates, many motion-sensitive neurons with a wide range of preferred directions respond to a stimulus moving in a single direction. These neurons are involved in direction perception, but it is not clear how perceptual decisions are related to the population response. We recorded the activities of MT neurons in rhesus monkeys while they discriminated closely related directions, and examined the relationship between the activities of neurons tuned to different directions and the monkeys' choices. Perceptual decisions were significantly correlated with the activities of the highest-precision neurons but not with those of the lowest-precision neurons. The combined performance of the high-precision neurons matched the monkeys' behavior, whereas the ability to predict behavior based on the entire active population was poor. These results suggest that fine discrimination decisions are crucially dependent on the activities of the most informative neurons.

NATURE | VOL 396 | 3 DECEMBER 1998 | www.nature.com that of the strobed segment (d s ) remains constant. The latency-difference hypothesis therefore predicts that the observed spatial lead of the moving central segment should increase.To test this prediction, we measured the spatial lead of the moving central segment as a function of the detectability of the central segment while keeping the detectability of the strobed segments constant. Here we use detectability to refer to the number of log units of luminance (Lu) above the detection threshold; detectability of the strobed segments was 0.3 Lu for subjects S.S.P. and G.P., and 0.5 Lu for T.L.N. The temporal lead of the moving central segment averaged across subjects increases systematically from 20 to 70 ms when its detectability increases by 1.0 Lu (Fig. 1b).Increasing the luminance of the strobed segments while keeping that of the moving central segment constant should decrease d s , while d m remains constant. The latencydifference hypothesis predicts that the observed spatial lead of the moving central segment should decrease and, if the luminance of the strobed segments is high enough, the moving central segment should be perceived to lag behind spatially. We tested this prediction by measuring spatial lead as a function of the detectability of the strobed segments, while keeping the detectability of the moving central segment constant (1.5 Lu above the detection threshold for subjects G.P. and T.L.N., and 0.8 Lu for S.S.P.). The observed temporal lead of the moving central segment averaged across subjects decreases systematically from 80 to ǁ30 ms as the detectability of the strobed segments increases by 1.5 to 2.0 Lu (Fig. 1c).These results support predictions of the latency-difference hypothesis and show that the motion-extrapolation mechanism does not compensate for stimulus-dependent variations in latency. Indeed, theoretical calculations show that the putative motionextrapolation mechanism must be undercompensating by at least 120 ms to account for the data in Fig. 1. But a motion-extrapolation mechanism that does not adequately compensate for variations in visual latency would not appreciably improve the accuracy of real-time visually guided behaviour.

The primary visual cortex (V1) receives its driving input from the eyes via the lateral geniculate nucleus (LGN) of the thalamus. The lateral pulvinar nucleus of the thalamus also projects to V1 but this input is little understood. We manipulated lateral pulvinar neural activity and assessed the effect on supra-granular layers of V1 that project to higher visual cortex. Reversibly inactivating lateral pulvinar prevented supra-granular V1 neurons from responding to visual stimulation. Reversible, focal excitation of lateral pulvinar receptive fields increased 4-fold the visual responses in coincident V1 receptive fields and shifted partially overlapping V1 receptive fields towards the center of excitation. V1 responses to regions surrounding the excited lateral pulvinar receptive fields were suppressed. LGN responses were unaffected by these lateral pulvinar manipulations. Excitation of lateral pulvinar after LGN lesion activated supra-granular layer V1 neurons. Thus, lateral pulvinar is able to powerfully control and gate information outflow from V1.

Fundamental to neuroscience is the understanding of how the language of neurons relates to behavior. In the lateral geniculate nucleus (LGN), cells show distinct properties such as selectivity for particular wavelengths, increments or decrements in contrast, or preference for fine detail versus rapid motion. No studies, however, have measured how LGN cells respond when an animal is challenged to make a perceptual decision using information within the receptive fields of those LGN cells. In this study we measured neural activity in the macaque LGN during a two-alternative, forced-choice (2AFC) contrast detection task or during a passive fixation task and found that a small proportion (13.5%) of single LGN parvocellular (P) and magnocellular (M) neurons matched the psychophysical performance of the monkey. The majority of LGN neurons measured in both tasks were not as sensitive as the monkey. The covariation between neural response and behavior (quantified as choice probability) was significantly above chance during active detection, even when there was no external stimulus. Interneuronal correlations and task-related gain modulations were negligible under the same condition. A bottom-up pooling model that used sensory neural responses to compute perceptual choices in the absence of interneuronal correlations could fully explain these results at the level of the LGN, supporting the hypothesis that the perceptual decision pool consists of multiple sensory neurons and that response fluctuations in these neurons can influence perception.

This paper introduces quantum neural networks (QNNs), a class of feedforward neural networks which are inherently capable of estimating the structure of a feature space in the form of fuzzy membership information. The hidden units of these networks develop quantized representations of the crisp sample information provided by the training set in various graded levels of certainty. Experimental results show that QNNs have an inherent ability for recognizing structures in the feature space that conventional feedforward neural networks with sigmoidal hidden units lack.