Figure 2. Mechanistic circuit model of attention effects.

(A) Schematic of an excitatory and inhibitory neuronal network model of attention (Huang et al., 2019) that extends the three-layer, spatially ordered network to include the orientation tuning and organization of V1. The network models the hierarchical connectivity between layer 4 of V1, layers 2 and 3 of V1, and V4. In this model, attention depolarizes the inhibitory neurons in V4 and increases the feedforward projection strength from layers 2 and 3 of V1 to V4. (B, C) We mapped the n-dimensional neuronal activity of our model to a two-dimensional space (a ring). Each dot represents the neuronal activity of the simulated population on a single trial and each color represents the trials for a given orientation. These fluctuations are more elongated in the (B) unattended state than in the (C) attended state. We then calculated the effects of these attentional changes on the performance of specific and general decoders (see Materials and methods). The axes are arbitrary units. (D–F) Comparisons of the modeled versus electrophysiologically recorded effects of attention on V4 population activity. (D) Firing rates of excitatory neurons increased, (E) mean correlated variability decreased, and (F) as illustrated with the first five largest eigenvalues of the shared component of the spike count covariance matrix from the V4 neurons, attention largely reduced the eigenvalue of the first mode. Attentional state denoted by marker color for the model (yellow: most attended; green: least attended) and electrophysiological data (yellow: cued; green: uncued). For the model: 30 samplings of n=50 neurons. Monkey 1 data illustrated for the electrophysiological data: n=46 days of recorded data. SEM error bars. Also see Figure 2—figure supplement 1.

Figure 2—figure supplement 1. The model reproduces the relationship between noise and signal correlations that is key to the general decoder hypothesis.

(A) As previously observed in electrophysiological data (Cohen and Maunsell, 2009; Cohen and Kohn, 2011), we observe a strong relationship between noise and signal correlations. During additional recordings collected during most recording sessions (for Monkey 1 illustrated here, n=37 days with additional recordings), the monkey was rewarded for passively fixating the center of the monitor while Gabors with randomly interleaved orientations were flashed at the receptive field location (‘Stim 2’ location in Figure 1C). The presented orientations spanned the full range of stimulus orientations (12 equally spaced orientations from 0° to 330°). We calculated the signal correlation for each pair of units based on their mean responses to each of the 12 orientations. We define the noise correlation for each pair of units as the average noise correlation for each orientation. The plot depicts signal correlation as a function of noise correlation across all recording sessions, binned into eight equally sized sets of unit pairs. Error bars represent SEM. (B) The model reproduces the relationship between noise and signal correlations. Signal correlation is plotted as a function of noise correlation, binned into 20 equally sized sets of unit pairs (n=2000 neurons), for each attentional modulation strength (green: least attended; yellow: most attended). The results were averaged over 50 tested orientations. (C) The slope of the relationship between noise and signal correlations (y-axis) decreases with increasing attentional modulation (x-axis). This suggests that noise is less aligned with signal correlation with increasing attentional modulation.