Perceptual decisions are explained by the accumulation of noisy evidence to a stopping bound.

a, Random dot motion discrimination task. The monkey fixates a central point. After a delay, two peripheral targets appear, followed by the random dot motion. When ready, the monkey reports the net direction of motion by making an eye movement to the corresponding target. Yellow shading indicates the response fields of a subset of neurons in area LIP that we refer to as Tin neurons (target in response field). b, Mean reaction times (top) and proportion of leftward choices (bottom) plotted as a function of motion strength and direction, indicated by the sign of the coherence: positive is leftward. Data (circles) are from all sessions from monkey M (black, 9684 trials) and monkey J (brown, 8142 trials). Solid lines are fits of a bounded drift-diffusion model. c, Drift-diffusion model. The decision process is depicted as a race between two accumulators: one integrating momentary evidence for left; the other for right. The momentary samples of evidence are sequential samples from a pair of negatively correlated (ρ = −0.71) Normal distributions. The decision is terminated when one accumulator reaches its positive bound. The example depicts leftward motion leading to a leftward decision.

Information about individual experimental sessions.

Population responses from LIP approximate drift-diffusion.

Rows show three types of population averages. The left columns show representative single-trial firing rates during the first 300 ms of evidence accumulation using two motion strengths: 0% and 25.6% coherence toward the left (contralateral) choice target. We highlight a number of trials with thick traces and these are the same trials in each of the rows. The right columns show the across-trial average responses for each coherence and direction. Motion strength and direction are indicated by color (legend) and aligned to motion onset (left) and saccade initiation (right). The gray bars under the motion-aligned averages indicate the 300 ms epoch used in the display of the single-trial responses (left panels). It begins when LIP first registers a signal related to the strength and direction of motion. Except for saccade-aligned response, trials are cut off 100 ms before saccade initiation. Error trials are excluded from the saccade-aligned averages, only. a, Ramp coding direction. The weight vector is established by regression to ramps from −1 to + 1 over the period of putative integration, from 200 ms after motion onset to 100 ms before saccade initiation (see Fig. S2). Only trials ending in left (contralateral) choices are used in the regression. b, First principal component (PC1) coding direction. c, Average firing rates of the subset of neurons that represent the left (contralateral) target. The weight vector consists of for each of the N neurons and 0 for all other neurons. Note the similarity of both the single-trial traces and the response averages produced by the different weighting strategies.

Variance and autocorrelation of the single trial signals.

The analyses here are based on samples of Sramp at six time points during the first 300 ms of putative integration, using all 0% and ±3.2% coherence trials (N = 5927). Samples are separated by the width of the boxcar filter (51 ms), beginning at t1 = 225 ms. The smoothed traces are detrended by the mean across trials and the value of Sramp(t0 = 0.2) on each trial. a, Variance increases as a function of time. The measure of variance is normalized so that it is 1 for the first sample. b, Autocorrelation of samples as a function of time and lag approximate the values expected from diffusion. The upper triangular portion of the 6 × 6 correlation matrix for unbounded diffusion (ri,j is represented by brightness). The values from the data (Sramp) are similar (right). c Nine of the 15 autocorrelation terms in b permit a more direct comparison of theory and data. The lower limb of the C-shaped function shows the decay in ri,j as a function of lag (ji). This is the top row of b. The upper limb shows the increase in ri,j as a function of time (for fixed lag). This is the lower diagonal in b. Note that the autocorrelations incorporate a free parameter, ф ≤ 1, that serves to correct for an unknown fraction of the measured variance that is not explained by diffusion (see Methods).

The population signal predictive of choice and RT is approximately one-dimensional.

Two binary decoders were trained to predict the choice (What-decoder) and its time (When-decoder) using the population responses in each session. The When-decoder predicts whether a saccadic response to the contralateral target will occur in the next 150 ms, but critically, its accuracy is evaluated based on its ability to predict choice. a, Choice decoding accuracy plotted as a function of time from motion onset (left) and time to saccadic choice (right). Values are averages across sessions. The What-decoder is either trained at the time point at which it is evaluated (time-dependent decoder, orange) or at the single time point indicated by the red arrow (t = 450 ms after motion onset; fixed training-time decoder, red). Both training procedures achieve high levels of accuracy. The When-decoder is trained to capture the time of response only on trials terminating with a left (contraversive) choice. The coding direction identified by this approach nonetheless predicts choice (green) nearly as well as the fixed training-time What-decoder. The black trace shows the accuracy of a What-decoder trained on simulated signals using a drift-diffusion model that approximates the behavioral data in Fig. 1. Shaded bar shows the epoch depicted in the next panel. b, The heat map shows the accuracy of a decoder trained at times along the abscissa and tested at times along the ordinate. Time is relative to motion onset (gray shading in a). In addition to training at t = 450 ms, the decoder can be trained at any time from 250 < t < 500 ms and achieve the same level of accuracy when tested at any single test time. The orange and red traces in a correspond to the main diagonal (x = y) and the column marked by the red arrow, respectively. c, Trial-averaged activity rendered by the projection of the population responses along the When coding direction, SWhen. Same conventions as in Fig. 2. d, Cosine similarity of five coding directions. The heatmap shows the mean values across the sessions, arranged like the lower triangular portion of a correlation matrix. The main diagonal is by definition 1. All but one of the 80 unique pairs (10 × 8 sessions) are greater than zero (p < 10−28, t-test, ℋ0: mean cosine similarity is 0). e, Correlation of single-trial diffusion traces. The Pearson correlations are calculated from ordered pairs of {Sx(ti), Sy(ti)}, where i is the sample time of the detrended signals. The detrending removes trial-averaged means for each signed coherence, leaving only the diffusion component.

The drift-diffusion signal is the decision variable.

The graphs show the leverage of single-trial drift-diffusion signals on choice and RT using only trials with RT ≥ 0.67 s. Rows correspond to the same coding directions as in Fig. 2. The graphs also demonstrate a reduction of the leverage of the samples at t ≤ 0.5 s by a later sample of the signal at t = 0.55 s. a, Leverage of single-trial drift-diffusion signals on choice. Leverage is the value of β1, the coefficient that multiplies S(t) in the GLM to predict choice (Eq. 7). The black traces show the increase in leverage as a function of time. Filled symbols show the leverage at t = 0.55 s. The green curve shows the leverage when the later sample is included in the GLM (mediation by later sample of the row signal). Open symbols show the leverage of at t = 0.55 s (same value as the filled symbol in bottom row). The red curves (top and middle rows) show the leverage of the signal when the later sample of is included in the GLM (mediation by the later sample of . b, Leverage of single-trial drift-diffusion signals on response time. Leverage is the correlation between S(t) and RT. Same conventions as in a; the green and red traces are partial correlation of S(t) and RT given Sramp, (top), along PC1 (SPC1, middle) and of neurons (, bottom). Black traces signify the unmediated relationship and green and red traces the leverage after accounting for activity at t = 0.55 ms of the signal itself (filled symbol) or STin activity (open symbol), respectively. Error bars signify s.e.m across sessions.

The representation of momentary evidence in area LIP.

a, Leftward preferring neurons. (left) Response to strong leftward (blue) and rightward (brown) motion during passive viewing. Traces are averages over neurons and trials. The neurons were selected for analysis based on this task, hence the stronger response to leftward is guaranteed. Note the short latency visual response to motion onset followed by the direction-selective (DS) response beginning ∼100 ms after motion onset. (right) Responses during decision aligned to motion onset and saccadic response. Response averages are grouped by direction and strength of motion (color legend). The neurons retain the same direction preference during passive viewing and decision-making. The responses are also graded as a function of motion strength. b, Rightward preferring neurons. Same conventions as a. c, Cumulative distribution of the times at which individual neurons start showing evidence-dependent activity. Evidence dependence emerges earlier in and neurons (purple) than neurons (green). Arrows indicate the mean onset of evidence-dependent activity in each signal. The markers at the end of the arrows show the s.e.m. across neurons. d, Correlation between the neural representation of motion evidence—the difference in activity of neurons selective for leftward and rightward motion —and the neural representation of the decision variable across different time points and lags. Horizontal and vertical lines indicate the onset of evidence-dependent activity in each signal. Positive correlations in the upper left triangle indicate that the decision variable at a time point is correlated with earlier activity of the evidence signal. e, Leverage of neural activity on choice for (green), (purple) and (yellow) neurons. f, Same as e for the correlation between neural activity and reaction time. The absence of negative correlation is explained by insufficient power (see Methods).

Effect of motion pulses on behavior (adapted from Stine et al., 2023).

a, Choice (bottom) and reaction time (top) data as a function of motion strength from two monkeys, plotted separately for trials with leftward (grey) and rightward (black) motion pulses. Positive motion strengths represent leftward motion. The pulses (100 ms) had a biasing effect equivalent to shifting the choice function left or right by ±1.4 % coh (p < .001, likelihood ratio test). b, Effect of motion pulses on choices as a function of time from the response. Pulses had a persistent effect on choices, consistent with temporal integration of motion evidence. Shading is ±1 s.e.

Derivation of a ramp coding-direction in neuronal state space.

Weights are assigned to each of the N simultaneously recorded neurons in each session using simple least squares regression to approximate a ramp from −1 to +1 on the interval from 200 ms after motion onset to 100 ms before saccade initiation. Only trials ending in left (contraversive) choices are included. The graph shows the quality of the regression on 6 trials from Session 1. Projection of the population firing rates on the vector of weights renders the single trial signal, Sramp(t).

Trial-averaged activity grouped by RT quantile

Rows show three types of population averages of single-trial responses. Choice and RT quantile are indicated by color (legend) and aligned to motion onset (left) and saccade initiation (right). a, Ramp coding direction. b, First principal component from PCA. c, Averages of the subset of neurons that have the left (contralateral) target in their response field. Correct and error trials are included in the both motion and saccade aligned averages.

Trial-averaged activity after subtracting the urgency component.

a, Ramp coding direction (Sramp)., b, PC1 (SPC1). direction, c, Mean activity of neurons . In all graphs, these are the same data as in Fig. 2, third column, after subtraction of the trace representing 0% coherence motion (gray).

Derivation of a When direction in neuronal state space.

Weights are assigned to each of the N simultaneously recorded neurons in each session using logistic regression to approximate a step that takes a value of 0 from 200 ms before motion onset to 150 ms before saccade initiation, and a value of 1 for the following 100 ms (the last 50 ms before the saccade are discarded). Only trials ending in left (contraversive) choices are included. The graph shows the quality of the regression on 8 example trials. The traces are shown from 200 ms after motion onset to 50 ms before the saccade. Projection of the population firing rates on the vector of weights renders the single trial signal, SWhen(t).

Single-trials and trial averaged signals furnished by the When- and What-decoders.

Same convention as Fig. 2, using the same single-trial examples.

The Tin neurons are not discoverable by their weight assignments.

The neurons contribute strongly and positively to all coding directions, but they are a small fraction of the sampled population. Here we ask whether the rank of a neuron’s weight might identify it as . The graphs are logistic fits

to the binary variable

Neurons with stronger positive weights are more likely to be , but even the largest weight percentile identifies a Tin neuron with a probability less than 0.4.

Cross-mediation of Single-trial correlations with behavior.

The figure extends the observations in Fig. 5 that a sample of at t = 550 ms after motion onset (i) reduces the leverage of earlier samples of on choice and RT (mediation of Sx on Sx) and (ii) also reduces the leverage of earlier samples of other signals, and (cross-mediation of Sx on Sy). The The heatmaps are 5 × 5 matrices of the mediation indices, ξCh and ξRT, (Eqs. 6 and 9), that is how a signal at t = 550ms mediates a signal at t = 400 ms after motion onset. The index is zero if there is no mediation by the later sample; one of the mediation is complete. The main diagonal (top left to bottom right) shows the mediation of Sx (0.55) on Sx (0.4). Values below the diagonal show cross-mediation of Sy by Sx; values above the diagonal show mediation of Sx by Sy. a, Leverage on choice. b, Leverage on RT. Notice that the matrices are not symmetric. For example, Sramp mediates the leverage of on choice more mediates the leverage of , and mediates the leverage of SWhen on RT more than SWhen mediates the leverage of .

Decoding choice from subsets of neurons.

Cross-validated, across-session mean accuracy of 7 choice decoders plotted as a function of time from motion onset (left) and time to saccadic choice (right). Solid traces indicate the accuracy of decoders trained at the time point at which they were evaluated (time-dependent). Dashed traces are generated by deriving decoder weights using only the activity at 450 ms from motion onset (red arrow, same as Fig. 4a) and applying these same weights at all other time points (fixed training-time). Trace colors indicate whether the decoders were trained using the activity of all neurons (gray), only and neurons (yellow), all but Tin neurons (purple), or all but Tin and Min neurons (blue). Decoding accuracy is greatly diminished without the contribution of these task-related neurons, which constitute only 28 ± 1.7% of the population (mean ± s.e.).