Neural basis for categorical boundaries in the primate pre-SMA during relative categorization of time intervals

Abstract

Perceptual categorization depends on the assignment of different stimuli to specific groups based, in principle, on the notion of flexible categorical boundaries. To determine the neural basis of categorical boundaries, we record the activity of pre-SMA neurons of monkeys executing an interval categorization task in which the limit between short and long categories changes between blocks of trials within a session. A large population of cells encodes this boundary by reaching a constant peak of activity close to the corresponding subjective limit. Notably, the time at which this peak is reached changes according to the categorical boundary of the current block, predicting the monkeys' categorical decision on a trial-by-trial basis. In addition, pre-SMA cells also represent the category selected by the monkeys and the outcome of the decision. These results suggest that the pre-SMA adaptively encodes subjective duration boundaries between short and long durations and contains crucial neural information to categorize intervals and evaluate the outcome of such perceptual decisions.

Fig. 1

Trial events during the interval categorization task and behavioral results. a Monkeys categorized the interval between the first and second stimulus presentation as “short” or “long” by moving the red cursor into the orange or blue circular targets, respectively. b An example of a possible target configuration, with the response circles 180° apart. Gray circles (not visible in the actual task) show the eight possible target locations. c The three blocks of stimuli employed in this study (T1, T2, and T3). The short/long implicit limits for T1, T2, and T3 are 350, 685, and 1195 ms, respectively. Note that some time durations could be correctly categorized either as “short” or as “long” depending on the particular block presented to the monkey. d Psychometric curves for the two monkeys for the three blocks of durations. Dots represent the mean (±SEM) probability of categorizing a particular interval as long. e Mean (±SEM) of the reaction times of the monkey 1 as a function of the test intervals for the three blocks of stimuli

Fig. 2

‘Boundary’ neurons encode the boundary between short and long categories across stimulus blocks. a Mean SDFs of a neuron for the test intervals of the T1 block. The two green vertical lines correspond to the two stimuli (s1 and s2) that define the test interval; the SDFs are aligned to the first stimulus. The time of peak activity (asterisk) occurs close to the implicit limit between short and long categories for this block (350 ms). The black vertical line indicates the target presentation (ta) and the gray rectangle the standard deviation of the movement initiation (rt). b Activity of the same neuron in a but recorded in the T3 block. The peak of activity shifted to the right, close to the limit for this block (1195 ms). c Histograms and corresponding Gaussian fittings of the mean times of peak activity for all boundary neurons across the three blocks (see the inset for color code). Arrowheads show the location of the implicit limit for each block. Asterisks indicate the mean times of peak activity of the neuron in a, b. d DL from both monkeys (red) and the standard deviation of the Gaussian functions in c as a function of the implicit limit for the three blocks. Lines show the best linear fits to the data. The linear increase in temporal variability (threshold) as a function of implicit limit duration (i.e., the scalar property of interval timing) is similar to the increase in the standard deviation of the peak time distributions for boundary neurons across the three blocks. The slope and constant are 54.5 and 0.09 for the psychometric behavior, and 45.2 and 0.06 for the boundary neuron distributions (both fits with p < 0.05). e Comparison between the behavioral constant error (mean ± SEM) and the time difference between the mean of the Gaussian functions in c and the implicit limit for the three blocks. Both the behavioral (red) and neural measures overestimate the shortest block (T1) and underestimate the longest block (T3)

Fig. 3

Iterative algorithm used to find the best regression model to describe the increase or decrease of instantaneous activity of boundary neurons over time with respect to a sensory event. a Raster plot and mean spikes density function (SDF; σ = 30 ms; black function) of the boundary cell in Fig. 2, for block 3 and the test interval duration of 1470 ms, aligned to the presentation of the first stimulus (s1). b A series of linear regression functions are displayed (blue lines), including the best model (thicker green line) identified by the algorithm. c Parameters extracted from the linear regression model for the identification of bisection neurons

Fig. 4

Trial-by-trial decoding of the monkey’s forthcoming perceptual decision employing the peak activity time of boundary neurons. a SDFs of the boundary neuron in Fig. 2a, b for trials in which the shortest test interval (870 ms, block 3) was presented to the monkey. Each color line represents an individual trial (eight trials are shown). Dots indicate the moment of the peak activity. s1 and s2: moments of presentation of the 1st and 2nd stimulus, respectively. Note that most activity peaks occur after the second stimulus. b Trial by trial activity of the same boundary neuron for one of the longer test intervals of the same block (1470 ms, block 3). Note that all activity peaks occur before the second stimulus. The interval between the peak activity (τ) and the second stimulus is indicated for one trial. c Number of errors in trial classification used as categorization cut criteria with a range from −500 to 500 ms in steps of 1 ms. The best limit criterion corresponds to the value of τ of the boundary neuron in a, b that minimized the classification error for trials as coming from actual short or long response trials. d Values of τ for the same cell as a function of the test intervals of the block 3 of the time categorization task. Note that for shorter intervals values were mainly positive, which implies that the peak activity occurred mainly after the presentation of the second stimulus. The opposite occurs for longer intervals. The red horizontal line corresponds to the best criterion in c. The color code is in the inset. e Psychometric (blue) and neurometric (red) functions. The neurometric curve was constructed from the trial probability below the best criterion as a function of test interval in d. Logistic functions were fitted to behavioral and neuronal data; the DLs were 112 ms for both curves and the PSE was 1052 ms for the behavior and 1105 ms for the neurometric function. Finally, the relation between the actual categorical behavior and the neural classification of the cell activity in c for the 96 trials was remarkably large (χ²-test = 63.8, P < 0.00001), as well as the mean boundary-choice probability (Methods section)

Fig. 5

Mean activity of the population of boundary neurons for the three blocks of the categorization task (T1, T2, and T3). The color code corresponds to the eight test intervals as indicated in the inset. The vertical line indicates the presentation of the first stimulus (s1). The vertical rectangle indicates the mean (+1 SD) time to peak activity from first stimulus computed from the Gaussian distributions of Fig. 2c

Fig. 6

A neural-network model of boundary cells generating categorical signals. a Center panel, the model consisted of an independent sensory and a boundary recurrent network whose outputs were read by a linear classifier. Each network consisted of 800 excitatory and 200 inhibitory neurons. Left panel, the sensory network received an input comprising of two pulses (s1 and s2) separated by a test interval, whereas the boundary network received a set of pulses with intervals coming from a Gaussian distribution, simulating boundary input responses. The pulses modulated the activity of each network cell through two input current types: AMPA, which induced facilitation, and GABAb, which corresponded to a slow inhibitory current. Right panel, PCA was used to project the profile of neural activation of the sensory and boundary networks onto three dimensions (X = Boundary PC1, Y = Boundary PC2, Z = Sensory PC1). A linear classifier (gray plane) was used to find the plane in the PCA space that could divide short and long intervals according to the psychometric performance of the monkey. Thus, the proportion of simulations (green ellipse for shorter interval and brown ellipse for long interval) that was to the right of the discrimination plane was considered short, while the proportion to the left was considered long. b Temporal pattern of activity of the sensory network population during the presentation of a short and a long interval (upper and lower panels, respectively). After the presentation of the second pulse, at 450 or 920 ms, the network presented a large phasic activation. The mean population firing rate is indicated in the middle panel. c Temporal pattern of activity of the boundary network population during the presentation of the same short and long intervals of (b) (upper and lower panels, respectively). When the longest interval was presented, the population of cells in the boundary network generated an up–down profile of activation. d Networkmetric curves generated by the model. The probability of long categorization was computed from the linear classifier in a right. e Constant error and f DL produced by the model for the three blocks of stimuli employed in the time categorization task

Fig. 7

Category-selective neurons. a Mean SDFs of a category-selective neuron of monkey 2 whose activity was modulated during the delay epoch by the future monkey’s categorical choice for short durations. The activity is segregated by the monkey’s choice (red: short, blue: long choice) and interval duration (indicated to the left of each graph). Same notation as in Fig. 2a. b Psychometric performance (green) of the monkey during the recording of the neuron in a, whose neurometric function is shown in orange. The choice probability of this cell was close to 1 and the χ²-test on the contingency table calculated between the decoded and the observed monkey’s choices across all trials was significant (x² = 54.43; p < 0. 001). c Population activity (mean ± SEM) of neurons of monkey 1 whose firing rate is significantly higher for short response trials during the delay epoch (red function). Orange asterisks indicate: (1) t-test significant differences in ten millisecond, non-overlapping bins (over the population SDF, short vs. long choices, 95% bootstrap confidence interval, 100 iterations). The single cell (a) and the population activity (c) are aligned to the presentation of the second stimulus (s2)

Fig. 8

Decision outcome neurons. a Example of a pre-SMA neuron with higher activity for incorrect and unrewarded trials during the intertrial interval. The activity was separated into correct (red) and incorrect (blue) trials. s1, s2, ta, and rt indicate the standard deviation for first stimulus, second stimulus, targets and reaction time, respectively. b Population activity (mean ± SEM) of the neurons of monkey 2 whose firing rate is significantly larger for incorrect trials (blue function). Orange asterisks indicate: (1) t-test significant differences in non-overlapping sliding windows (over the population SDF, correct vs. incorrect, 95% bootstrap confidence interval, 100 iterations). The single cell (a) and the population activity (b) are aligned to the expected time of reward (re, red vertical line). c Cumulative distributions of the response onset latencies of the neurons selective for correct (orange) and incorrect (green) categorical decision outcomes during de intertrial interval. Note that the response onset latencies are shifted to the right for incorrect outcomes and are significantly larger for correct trials (Kolmogorov–Smirnov test, p < 0.0001)

Fig. 9

Neural activity encoding boundary, category, and trial outcome emerged sequentially throughout the task. Times at which pre-SMA neurons of monkey 2 carried significant boundary, category or outcome information for T1 (a), T2 (b), and T3. c Each row corresponds to a single neuron. Orange asterisk in a left indicates the neuron in Fig. 7a. Blue asterisk indicates neuron in Fig. 8a. Arrows at the bottom of each graph indicate task events. Boundary-related activity (left panel, brown) is aligned to the interval onset (s1); the offset of the shortest (s2 short) and longest (s2 long) intervals are indicated. Category-related activity (center panel, green) is aligned to the interval offset (s2). Outcome-related activity (right panel, blue) is aligned to reward delivery (re)