Fig 1.
Hypotheses on the sources of choice correlations in sensory area.
(A) Joint activity of the population. The point cloud represents neuronal activities colored by stimulus direction. The neural space can be divided into stimulus and non-stimulus axes. (B) Noise correlation is any elongation of the joint activity point cloud for repeats of the same stimulus. (C) Optimal readout. The optimal decision boundary is a criterion line orthogonal to the stimulus axis. All CP is due to readout and there is no CP in the non-stimulus axis. (D) Suboptimal readout. The decision boundary is not orthogonal to the stimulus axis. CP exist in both axes. (E) Corrupting feedback. The choice is fed back and pushes variability along the stimulus axis. This increases CP along the stimulus axis without affecting the non-stimulus axis, and causes more variability along the stimulus axis. (F) Non-corrupting feedback. Feedback pushes choice information in the non-stimulus axis and increases CP in the non-stimulus axis without adding CP in the stimulus axis.
Fig 2.
Experimental setup: Motion discrimination task.
Trials started with a fixation at the center of the monitor. 7 consecutive motion pulses were presented to the monkey while monkeys hold the fixation. Each motion pulse consists of drifting and flickering Gabor patches and lasts 150 ms. Random, signed motion strength was determined by changing the proportion of drifting vs. flickering Gabors patches. Monkeys reported their choice depending on the net direction by making a saccade to one of the two choice targets after the fixation point to disappear.
Fig 3.
Probabilistic description of a single trial using variational latent Gaussian process method and resulting noise correlation.
(A; top) Simultaneously recorded spike trains of the MT units in an example trial aligned to stimulus onset (yt in Eq 2). (A; bottom) Corresponding 4-dimensional factors. The rank-4 matrix multiplication of the loading matrix (matrix A in Eq (2)), and latent factors are exponentiated to produce the population rate. The loading matrix is rotated to maximize stimulus encoding (see Fig 4), so that the first column has the strongest stimulus response. The inferred latent factors (xt in Eq (2)) are colored to indicate the respective factors corresponding to the loading matrix. (B) The pairwise noise correlation matrices (neuron by neuron) for the sessions with frozen trials (trials with identical stimulus). The lower triangles are the correlations calculated from the raw data, and the upper triangles are the correlations from the reconstruction by the inferred 4-dimensional latent factors. Time bin size 100 ms.
Fig 4.
Visual motion pulse information encoded in one dimension.
(A) Pulse-triggered average of three example sessions, one from each monkey, are shown. The factors are rotated such that most of the stimulus power is in the first factor. They visualize the weights of pulses on the latent factors that were estimated from respective sessions. The color gradient indicates the seven pulses of visual motion stimuli. Each pulse last 150ms. (B) The power of each factor that explains the variation contributed by the stimuli to the factors. Each marker indicates one session, the shape indicates the animal and the color indicate the respective example sessions in (A).
Fig 5.
Data analysis pipeline and nested model comparison.
(1) Extract latent factors. (2) Align latent factors to stimulus & null dimensions. (3) Map dimensions of latent factors into real-valued scalars. (4) Pool the choice-mapping over all sessions and perform nested log-likelihood tests.
Fig 6.
Choice probabilities of latent factors for each monkey.
Contours corresponds to 50%, 90%, 99% quantities of the choice-mapped stimulus and non-stimulus trial distribution. The IN choice distribution (red-shade contours) is biased upward, indicating existence of the choice information in the non-stimulus axes. The pooled choice probability estimated using the choice-mapped stimulus-axis, non-stimulus-axes (the 3-dimensional subspace orthogonal to the stimulus-axis), and all 4 dimensions of the latent factors are 0.546, 0.591, and 0.621 respectively. The estimated population spike count choice probability is 0.627. For nested statistical tests of the corresponding regression models, see main text and Fig 5.
Fig 7.
Time course of choice probability in the latent stimulus subspace, latent non-stimulus subspace and neuron population space suggests feedback from the decision-making process.
Decoders were fit to early (yellow), middle (red), and late (purple) periods (300 ms, marked by the colored bars) of non-stimulus latent factors to predict choice. We used the resulting weights of the decoders to perform choice-mapping on the whole time interval divided into 100 ms non-overlapping moving windows (aligned at the center). The colored curves correspond to the choice probability time course using the respective decoder.
Fig 8.
Four possible sources of neural-choice correlation.
1-dimensional stimulus drive to MT is picked up as population variability along with other noise correlations denoted x1(t), x2(t), x3(t). To optimally perform the task, the choice should rely on only the stimulus dimension, and hence noise in x1 shows up as CP in relevant units reflecting their ‘readout’ strategy (case 1). Non-optimal readout can provide CP through stimulus-irrelevant variability (case 3). Alternatively, feedback from the decision-making process to MT can provide choice-correlation in the stimulus-irrelevant subspace (case 4) without corrupting the optimal representation or the stimulus driven shared dimension (case 2) causing non-optimal behavior.