Temporal integration is a robust feature of perceptual decisions
Abstract
Making informed decisions in noisy environments requires integrating sensory information over time. However, recent work has suggested that it may be difficult to determine whether an animal’s decisionmaking strategy relies on evidence integration or not. In particular, strategies based on extremadetection or random snapshots of the evidence stream may be difficult or even impossible to distinguish from classic evidence integration. Moreover, such nonintegration strategies might be surprisingly common in experiments that aimed to study decisions based on integration. To determine whether temporal integration is central to perceptual decisionmaking, we developed a new modelbased approach for comparing temporal integration against alternative ‘nonintegration’ strategies for tasks in which the sensory signal is composed of discrete stimulus samples. We applied these methods to behavioral data from monkeys, rats, and humans performing a variety of sensory decisionmaking tasks. In all species and tasks, we found converging evidence in favor of temporal integration. First, in all observers across studies, the integration model better accounted for standard behavioral statistics such as psychometric curves and psychophysical kernels. Second, we found that sensory samples with large evidence do not contribute disproportionately to subject choices, as predicted by an extremadetection strategy. Finally, we provide a direct confirmation of temporal integration by showing that the sum of both early and late evidence contributed to observer decisions. Overall, our results provide experimental evidence suggesting that temporal integration is an ubiquitous feature in mammalian perceptual decisionmaking. Our study also highlights the benefits of using experimental paradigms where the temporal stream of sensory evidence is controlled explicitly by the experimenter, and known precisely by the analyst, to characterize the temporal properties of the decision process.
Editor's evaluation
This manuscript tests an important assumption about how sensory information is processed and used to guide motor choices. The widely held assumption is that sensorymotor circuits are capable of integrating evidence, but the validity and generality of this 'principle' have been recently questioned by studies suggesting that other computational operations may lead to similar psychophysical results, mimicking integration without actually performing it. This study makes a compelling case that the integration assumption was likely correct all along and that the model mimicry can be easily disambiguated by using appropriate sensory stimuli and task designs that permit rigorous analyses.
https://doi.org/10.7554/eLife.84045.sa0Introduction
Perceptual decisionmaking is thought to rely on the temporal integration of noisy sensory information on a timescale of hundreds of milliseconds to seconds. Temporal integration corresponds to summing over time the evidence provided by each new sensory stimulus, and optimizes perceptual judgments in face of noise (Bogacz et al., 2006; Gold and Shadlen, 2007). A perceptual decision can then be made on the basis of this accumulated evidence, either as some threshold on accumulated evidence is reached, or if some internal or external cue signals the need to initiate a response.
Although many behavioral and neural results are consistent with this integration framework, temporal integration is a feature that has often been taken for granted rather than explicitly tested. Recently, the claim that standard perceptual decisionmaking tasks rely on (or even frequently elicit) temporal integration has been challenged by theoretical results showing that nonintegration strategies can produce behavior that carries superficial signatures of temporal integration (Stine et al., 2020). These signatures include the relationship between stimulus difficulty, stimulus duration, and behavioral accuracy, the precise temporal weighting of sensory information on the decisions, and the patterns of reaction times.
Here, we propose new analytical tools for directly assessing integration and nonintegration strategies from fixed or variableduration paradigms where, critically, the experimenter controls the fluctuations in perceptual evidence over time within each trial (discretesample stimulus, or DSS). By leveraging these controlled fluctuations, our methods allow us to make direct comparisons between integration and nonintegration strategies. We apply these tools to assess temporal integration in data from monkeys, humans, and rats that performed a variety of perceptual decisionmaking tasks with DSS. Applying these analyses to these behavioral datasets yields strong evidence that perceptual decisionmaking tasks in all three species rely on temporal integration. Temporal integration, a critical element of many major theories of perception at both the neural and behavioral levels, is indeed a robust and pervasive aspect of mammalian behavior. Our results also illuminate the power of targeted stimulus design and statistical analysis to test specific features of behavior.
Results
Integration and nonintegration models
In a typical perceptual evidenceintegration experiment (Figure 1A), an observer is presented in each trial with a timevarying stimulus and must report which of two possible stimulus categories it belongs to. Typical examples include judging whether a dynamic visual stimulus is moving leftwards or rightwards (Yates et al., 2017; Katz et al., 2015); whether the orientation of a set of gratings is more aligned with cardinal or diagonal directions (Wyart et al., 2012) whether a combination of tones is dominated by high or low frequencies (Morillon et al., 2014; HermosoMendizabal et al., 2020; Znamenskiy and Zador, 2013); which of two acoustic streams is more intense or dense (Brunton et al., 2013; PardoVazquez et al., 2019; Cisek et al., 2009). Such paradigms have been used extensively in humans, nonhuman primates, and rodents. Here, we focus on experiments in which observers report their choice at the end of a period whose duration is controlled by the experimenter (Kiani and Shadlen, 2009; Wyart et al., 2012; Brunton et al., 2013; Raposo et al., 2012), in contrast to socalled ‘reaction time’ tasks, in which the observer can respond after viewing as brief a portion of the stimulus as they wish (Roitman and Shadlen, 2002; Znamenskiy and Zador, 2013; PardoVazquez et al., 2019; HermosoMendizabal et al., 2020).
Moreover, we focus on experimental paradigms in which the sensory evidence in favor of each category arrives in a sequence of discrete samples. Samples can correspond to motion pulses (Yates et al., 2017), individual gratings (Wyart et al., 2012), acoustic tones (Morillon et al., 2014; HermosoMendizabal et al., 2020; Znamenskiy and Zador, 2013), numbers (Bronfman et al., 2015; Cisek et al., 2009), or symbols representing category probabilities (Yang and Shadlen, 2007). We refer to this configuration as the DSS paradigm. In this paradigm, the perceptual evidence provided by each sample can be controlled independently, allowing for detailed analyses of how different samples contribute to the behavioral response. The DSS framework can be contrasted with experiments in which the experimenter specifies only the mean stimulus strength on each trial, and variations in sensory evidence over time are not finely controlled or are not easily determined from the raw spatiotemporal stimulus.
Tasks using the DSS paradigm are classically thought to rely on sequential accumulation of the stimulus evidence (Bogacz et al., 2006), which we refer to here as temporal integration. Figure 1A shows an example stimulus sequence composed of n samples that provide differing amounts of evidence in favor of one alternative vs. another (‘A’ vs. ‘B’). In the temporal integration model, the accumulated evidence fluctuates as new samples are integrated and finishes at a positive value indicating overall evidence for stimulus category A (Figure 1B). This integration process can be formalized by defining the decision variable or accumulated evidence ${x}_{i}$ and its updating dynamics across stimulus samples: ${x}_{i}={x}_{i1}+{m}_{i}$ where ${{m}_{i}=S}_{i}+{\epsilon}_{i}$ represents a noisy version of the true stimulus evidence ${S}_{i}$ in the ith sample corrupted by sensory noise ${\epsilon}_{i}$ . The binary decision r is simply based on the sign of the accumulated evidence ${x}_{n}$ at the end of the sample sequence (composed of n samples): $r=A$ if ${x}_{n}>0$, and $r=B$ if ${x}_{n}<0$. This procedure corresponds to the normative strategy with uniform weighting that maximizes accuracy. For such perfect integration, ${x}_{n}={\Sigma}_{i}{S}_{i}+{\Sigma}_{i}{\epsilon}_{i}$ , so that the probability of response A is $p(r=A)=\Phi \left({\Sigma}_{i}\beta {S}_{i}\right)$ where $\Phi $ is the cumulative normal distribution function (the normative weight for the stimuli β depends on the noise variance $Var\left(\epsilon \right)$ and the number of samples through $\beta =1/\sqrt{nVar\left(\epsilon \right)}$). Departures from optimality in the accumulation process such as accumulation leak, categorization dynamics, divisive normalization, sensory adaptation, or sticky boundaries may however yield unequal weighting of the different samples (Yates et al., 2017; Brunton et al., 2013; PratOrtega et al., 2021; Bronfman et al., 2016; Keung et al., 2019; Keung et al., 2020). To accommodate for these, we allowed the integration model to take any arbitrary weighting of the samples: $p(r=A)=\Phi \left({{\beta}_{0}+\Sigma}_{i}{\beta}_{i}{S}_{i}\right)$ (see Methods for details). The mapping from final accumulated evidence to choice was probabilistic, to account for the effects of noise from different sources in the decisionmaking process (Drugowitsch et al., 2016). Thus, this model represented an approximate statistical description for any generative model relying on temporal integration of the stimulus evidence.
Although it has been commonly assumed that observers use evidenceintegration strategies to perform these psychophysical tasks, recent work has suggested that observers may employ nonintegration strategies instead (Stine et al., 2020). Here, we consider two specific alternative models. The first nonintegration model corresponds to an extremadetection model (Waskom and Kiani, 2018; Stine et al., 2020; Ditterich, 2006). In this model, observers do not integrate evidence across samples but instead base their decision on extreme or salient bits of evidence. More specifically, the observer commits to a decision based on the first sample i in the stimulus sequence that exceeds one of the two symmetrical thresholds, that is such that $\left{m}_{i}\right\ge \theta $. In our example stimulus, the first sample that reaches this threshold in evidence space is the fifth sample, which points toward stimulus category B, so response B is selected (Figure 1C). This policy can be viewed as a memoryless decision process with sticky bounds. If the stimulus sequence contains no extreme samples, so that neither threshold is reached, the observer selects a response at random. We also explored an alternative mechanism where in such cases the response is based on the last sample in the sequence, following Stine et al., 2020; and a variant of the model where the decision threshold is different on every sample position.
The second nonintegration model corresponds to the snapshot model (Stine et al., 2020; Pinto et al., 2018). In this model, the observer attends to only one sample i within the stimulus sequence, and makes a decision based solely on the evidence from the attended sample: $r=A$ if ${m}_{i}>0$, and $r=B$ if ${m}_{i}<0$. The position in the sequence of the attended sample is randomly selected on each trial. In our example, the fourth sample is randomly selected, and since it contains evidence toward stimulus category A, response A is selected (Figure 1D). We considered variants of this model that gave it additional flexibility, including: allowing the prior probability over the attended sample to depend on its position in the sequence using a nonparametric probability mass function estimated from the data; allowing for deterministic vs. probabilistic decisionmaking rule based on the attended evidence; including attentional lapses that were either fixed to 0.02 (split equally between leftward and rightward responses) or estimated from behavioral data. We finally considered a variant of the snapshot model where the decision was made based on a subsequence of K consecutive samples within the main stimulus sequence ($1\le K<n$), rather than based on a single sample.
Standard behavioral statistics favor integration accounts of pulsebased motion perception in primates
To compare the three decisionmaking models defined above (i.e., temporal integration, extremadetection, snapshots), we first examined behavioral data from two monkeys performing a fixedduration motion integration task (Yates et al., 2017). In this experiment, each stimulus was composed of a sequence of 7 motion samples of 150 ms each where the motion strength toward left or right was manipulated independently for each sample. At the end of the stimulus sequence, monkeys reported with a saccade whether the overall sequence contained more motion toward the left or right direction. The animals performed 72,137 and 33,416 trials for monkey N and monkey P, respectively, allowing for indepth dissection of their response patterns.
We fit the three models (and their variants) to the responses for each animal individually (see Figure 2—figure supplement 1 for estimated parameters for the different models). We then simulated the fitted model and computed, for simulated and experimental data, the psychophysical kernels capturing the weights of the different sensory samples based on their position in the stimulus sequence (Figure 2B). Psychophysical kernels were nonmonotonic and differed in shape between the two animals, probably reflecting the complex contributions of various dynamics and suboptimalities along the sensory and decision pathways (Yates et al., 2017; Levi and Huk, 2020).
The temporal profile of the kernel was perfectly matched by the integration model, almost by design, as we gave full flexibility to the model to adjust the sample weights. The snapshot model was provided with similar flexibility, as the prior probability of attending each sample could be fully adjusted to the monkey decisions. However, the snapshot model could not match the experimental psychophysical kernel as accurately. It consistently underestimated the magnitude of weighting in monkey P (Figure 2B, bottom row). The extremadetection model was not endowed with such flexibility of sensory weighting. On the contrary, since the decision was based on the first sample in the sequence reaching a certain criterion, this inevitably generates a primacy effect in the psychophysical kernels – or at best a flat weighting (Stine et al., 2020). The model thus failed to capture the nonmonotonic psychophysical kernels from animal data.
Next, we looked at the psychometric curves and choice accuracy predictions of each fitted model (Figure 2C, D). Stine et al., 2020 have argued that integration and nonintegration models can capture the psychometric curves equally well. For both animals, the accuracy and psychometric curves were accurately captured by the integration model. In line with Stine et al., we also found that both nonintegration models could reproduce the shape of the psychometric curve in monkey N, although the quantitative fit was always better for the integration than nonintegration models. By contrast both nonintegration models failed to capture the psychometric curve for monkey P (Figure 2B, bottom row). More systematically, the overall accuracy, which is an aggregate measure of the psychometric curve, clearly differed between models, as the accuracy of the nonintegration models systematically deviated from animal data for both animals (Figure 2C). In other words, all models produce the same type of psychometric curves up to a scaling factor, and this scaling factor (directly linked to the model accuracy) is key to differentiate model fits. For the snapshot model in monkey P, this discrepancy was explained because the model, limited to using one stimulus sample, could not reach the performance of the animal (compare the maximum accuracy of the model indicated by the blue mark with the accuracy of the animal). (This also explains why the psychophysical kernel of the snapshot model underestimated the true kernel in monkey P.) For the extremadetection model in monkey P and for both nonintegration models in the other animal (monkey N), the model accuracy is not bounded below the subject’s accuracy. In such cases, the model can produce betterthanobserved accuracy for certain parameter ranges, but these are not the parameters found by the maximum likelihood procedure, probably because they produce a pattern of errors that is inconsistent with the observed pattern of errors. This indicates an inability of the models to match the pattern of errors of the animal (see Discussion).
Finally, we assessed quantitatively which model provided the best fit, while correcting for model complexity using the Akaike information criterion (AIC, Figure 2A). In both monkeys, AIC favored the integration model over the two nonintegration models by a very large margin. We also explored whether (previously unpublished) elaborations of the extremadetection and snapshot models could provide a better match to the behavioral metrics considered above (Figure 2—figure supplements 2 and 3). We found using the AIC metric that the integration model was preferred over all variants of both nonintegration models, for both monkeys. Moreover, these model variants could not replicate the psychophysical kernels as well as the integration model did (Figure 2—figure supplements 2 and 3).
In conclusion, while psychometric curves may not always discriminate between integration and nonintegration strategies, other metrics including psychophysical kernels, predicted accuracy and quality of fit (AIC) support temporal integration in monkey perceptual decisions. For one model in one monkey (the snapshot model in monkey P), even the simple metric of overall accuracy compellingly supported temporal integration (Figure 2C). For the other monkey and/or model, where the distinction was less clear, our modelbased approach allowed us to leverage these other metrics to reveal strong support for the temporal integration model (Figure 2A–C). While these data rely only on two experimental subjects, we show below further evidence supporting the integration model in humans and rats.
Temporal integration is more likely than the extremadetection model: evidence from a unique subset of trials
While formal model comparison leads us to reject the nonintegration models in favor of the integration models, it is informative to examine qualitative features of the animal strategies and identify how nonintegration models failed to capture them. We started by designing two analyses aimed at testing whether choices were consistent with the extremadetection model, namely by testing whether choices were strongly correlated with the largest evidence samples. In the first analysis, we looked at the subset of trials where the evidence provided by the largest evidence sample in the sequence was at odds with the total evidence in the sequence: we show one example in Figure 3B, where the largest evidence sample points toward response B, while the overall evidence points toward response A. These ‘disagree trials’ represent a substantial minority of the whole dataset: 1865 trials (2.6%) in monkey N and 1831 trials (5.5%) in monkey P. If integration is present, the response of the animal should in general be aligned with the total evidence from the sequence (Figure 3A, red bars). By contrast, if it followed the extremadetection model (Figure 1C), it should in general follow the largest evidence sample (Figure 3A, green bars). In both monkeys, animal choices were more often than not aligned with the integrated evidence (Figure 3A, black bars), as predicted by the integration model. The responses generated from the extremadetection model tended to align more with the largest evidence sample, although that behavior was somehow erratic (for monkey N) due to the large estimated decision noise in the model. This rules out that monkey decisions rely on a memoryless strategy of simply detecting large evidence samples, discarding all information provided by lower evidence samples. Our results complement a previous analysis on disagree trials in this task (Levi et al., 2018), by explicitly comparing monkey behavior to model predictions.
We reasoned that the extremadetection would also leave a clear signature in the ‘subjective weight’ of the samples, defined as the impact of each sample on the decision as a function of absolute sample evidence (Yang and Shadlen, 2007; Waskom and Kiani, 2018; Nienborg and Cumming, 2007). The extremadetection model predicts that, in principle, samples whose evidence is below the threshold have little impact on the decision, while samples whose evidence is above the threshold have full impact on the decision. By contrast, the integration model predicts that subjective weight should grow linearly with sample evidence. We estimated subjective weights from monkey choices using a regression method similar in spirit to previous methods (Yang and Shadlen, 2007; Waskom and Kiani, 2018), taking the form $p\left({r}_{t}=A\right)\text{}=\text{}\sigma \left({\beta}_{0}+{\mathrm{\Sigma}}_{i\in \left[\mathrm{1...}n\right]}{\beta}_{i}\text{}f\left({S}_{t,i}\right)\right)$, where σ is a sigmoidal function. Here f is a function that captures the subjective weight of the sample as a function of its associated evidence. Whereas previous methods estimated subjective weights assuming a uniform psychophysical kernel, our method estimated simultaneously subjective weights $f\left(S\right)$ and the psychophysical kernel $\beta $, thus removing potential estimation biases due to unequal weighting of sample evidence (see Methods). In both monkeys, we indeed found that the subjective weight depends linearly on sample evidence for low to median values of sample evidence (motion pulse lower than 6), in agreement with the integration model (Figure 3—figure supplement 1). Counter to our predictions, simulated data of the extremadetection model displayed the same linear pattern for low to median values of sample evidence. We realized this was due to the very high estimated sensory noise (Figure 2—figure supplement 1), such that, according to the model, even samples with minimal sample evidence were likely to reach the extremadetection threshold. In other words, unlike the previous analyses, inferring the subjective weights used by animals was inconclusive as to whether animals deployed the extremadetection strategy. This somewhat surprising dependency reinforces the importance of validating intuitions by fitting and simulating models (Wilson and Collins, 2019).
Impact of early and late stimulus evidence onto choice shows direct evidence for temporal integration
Following model comparisons favoring integration over both snapshot and extremadetection models, the immediately previous analysis relied on a special subset of trials to provide an additional, and perhaps more intuitive, signature of integration, which ruled out extremadetection as a possible strategy of either monkey. We next employed another novel analysis specifically designed to tease apart unique signatures of the integration and snapshot models. More specifically, we tested whether decisions were based on the information from only one part of the sequence, as predicted by the snapshot model, or from the full sequence, as predicted by the integration model. To facilitate the analysis, we defined early evidence E_{t} by grouping evidence from the first three samples in the sequence, and late evidence L_{t}, as the grouped evidence from the last four samples. We then displayed the proportion of rightward responses as a function of both early and late evidence in a graphical representation that we call integration map (Figure 4A). A pure integration strategy corresponds to summing early and late evidence equally, which can be formalized as $p\left(r\right)=\sigma ({E}_{t}+{L}_{t})$, where $\sigma $ is a sigmoidal function. Because this only depends on the sum ${E}_{t}+{L}_{t}$ , the probability of response is invariant to changes in the $({E}_{t},{L}_{t})$ space along the diagonals, which leave the sum unchanged. These diagonals correspond to isolines of the integration map (Figure 4A, left; Figure 4—figure supplement 2A). In other words, straight diagonal isolines in the integration map reflect the fact that the decision only depends on the sum of evidence ${E}_{t}+{L}_{t}$ . Straight isolines thus constitute a specific signature of evidence integration.
We contrasted this integration map with the one obtained from a nonintegration strategy (Figure 4A, middle panel; Figure 4—figure supplement 2B). There we assumed that the decision depends either on the early evidence or on the late evidence, as in the snapshot model, with equal probability. This can be formalized as $p\left(r\right)=0.5\sigma \left({E}_{t}\right)+0.5\sigma \left({L}_{t}\right)$. In this case, if late evidence is null ($\sigma \left({L}_{t}\right)=0.5$) and early evidence is very strong toward the right ($\sigma \left({E}_{t}\right)\simeq 1$) the overall probability for rightward response is $p\left(r\right)=0.75$. This probability contrasts with that obtained in the integration case where the early evidence would dominate and lead to an overwhelming proportion of rightward responses, that is $p\left(r\right)\simeq 1$. The 25% of leftwards responses yielded by the nonintegration model correspond to trials where only the late (uninformative) part of the stimulus is attended and a random response to the left is drawn. More generally, in regions of the space in which either early or late evidence take large absolute values, their corresponding probability of choice saturates to 0 or 1, when that evidence is attended, so the overall response probability becomes only sensitive to the other evidence. As a result, the equiprobable lines bend toward the horizontal and vertical axes (Figure 4A, middle). Finally, to compare predictions from both integration and nonintegration models to monkey behavior, we plotted the integration maps for both monkeys (Figure 4A, right; Figure 4—figure supplement 1A). The isolines were almost straight diagonal lines and showed no consistent curvature toward the horizontal and vertical axes. This provides direct evidence that monkey responses predominantly depend on the sum of early and late evidence – a clear signature of temporal integration.
We derived subsequent tests based on the integration map. We computed conditional psychometric curves as the probability for rightward responses as a function of early evidence ${E}_{t}$ , conditioned on late evidence value ${L}_{t}$ (Figure 4B; Figure 4—figure supplement 1B). From the integration formula $p\left(r\right)=\sigma ({E}_{t}+{L}_{t})$, we see that a change in late evidence value corresponds to a horizontal shift of the conditional psychometric curves. By contrast, according to the nonintegration formula $p\left(r\right)=0.5\sigma \left({E}_{t}\right)+0.5\sigma \left({L}_{t}\right)$, conditioning on different values of late evidence adds a fixed value to the response probability irrespective of early evidence, a vertical shift akin to that introduced by lapse responses (Figure 4B, middle panel). The conditional psychometric curves for monkeys (Figure 4B, right panel; Figure 4—figure supplement 1 and Figure 4—figure supplement 2) displayed horizontal shifts as late evidence was changed, consistently with the integration hypothesis. We sought to quantify these shifts in better detail. To this purpose, we fitted each conditional psychometric curve with the formula $p\left(r\right)=(1{\pi}_{L}{\pi}_{R})\sigma (\alpha {E}_{t}+\beta )+{\pi}_{R}$ , where ${\pi}_{L}$ , ${\pi}_{R}$ , $\alpha $, and $\beta $ correspond to the left lapse, right lapse, sensitivity, and lateral bias parameters, respectively (Figure 4C, Figure 4—figure supplement 1 and Figure 4—figure supplement 2). The integration model predicts that the bias parameter $\beta $ should vary linearly with L_{t}, while lapse parameters should remain null (Figure 4D, left panel). By contrast, the nonintegration model predicts that the horizontal shift parameter $\beta $ should remain constant while left and right lapse parameters $({\pi}_{L},{\pi}_{R})$ should vary (middle panel), as these lapse parameters correspond to the trials where early evidence is not attended and the response depends simply on late evidence. Both monkeys showed a very strong linear dependence between late evidence and the horizontal shift $\beta $ (Figure 4D, right panel; see also Figure 4—figure supplement 1), further supporting that late evidence is summed to early evidence. By contrast, the lapse parameters showed no consistent relationship with late evidence ${L}_{t}$ (Figure 4E, right panel). Finally, we directly assessed the similarities between the integration maps from monkey responses and from simulated responses for the three models (integration, snapshot, and extremadetection). The modeldata correlation was larger in the integration model than in the nonintegration strategies for both monkeys (Figure 4E; unpaired ttest on bootstrapped r values: p < 0.001 for each animal and comparison against extremadetection and against snapshot model). Overall, integration maps allow to dissect how early and late parts of the stimulus sequence are combined to produce a behavioral response. In both monkeys, these maps carried signatures of temporal integration. For monkey N, the integration model and the data look very similar (Figure 4—figure supplement 2). For monkey P, there is still a qualitative dependency that deviates from nonintegration, but which is not as uniquely matched to the integration strategy (although the imperfect coverage of the twodimensional space impedes further investigations; Figure 4—figure supplement 1). Thus, complementing the statistical model tests favoring integration, this richer visualization allows the data to show us that some degree of integration is occurring, albeit not perfect.
Visual orientation discrimination in humans relies on temporal integration
Overall, all our analyses converged to support the idea that monkey decisions in a fixedduration motion discrimination task relied on temporal integration. We explored whether the same results would hold for two other species and perceptual paradigms. We first analyzed the behavioral responses from nine human subjects performing a variableduration orientation discrimination task (Cheadle et al., 2014). In each trial, a sequence of 5–10 gratings with a certain orientation were shown to the subject, and the subject had to report whether they thought the gratings were overall mostly aligned to the left or to the right diagonal. In this task, the experimenter can control the evidence provided by each sample by adjusting the orientation of the grating. We performed the same analyses on the participant responses as on monkey data. As for monkeys, we found that the integration model nicely captured psychometric curves, participant accuracy and psychophysical kernels (Figure 5A–C, red curves and symbols). By contrast, both nonintegration models failed to capture these patterns (Figure 5A–C, blue and green curves and symbols). The accuracy from both models consistently underestimated participant performance: eight and six out of nine subjects outperformed the maximum performance for the snapshot and extremadetection models, respectively (Figure 5—figure supplement 1). This suggests that human participants achieved such accuracy by integrating sensory evidence over successive samples. Moreover, subjects overall weighted more later samples (Figure 5C), which is inconsistent with the extremadetection mechanism. A formal model comparison confirmed that in each participant, the integration model provided a far better account of subject responses than either of the nonintegration models did (Figure 5D). We then assessed how subjects combined information from weak and strong evidence samples into their decisions, using the same analyses as for monkeys. As predicted by the integration model, but not by the extremadetection model, human choices consistently aligned with the total stimulus evidence and not simply with the strongest evidence sample (Figure 5E). Finally, the average integration map for early and late evidence within the stimulus sequence displayed nearly linear diagonal isolines, showing that both were integrated into the response (Figure 5F). Integration maps from participants correlated better with maps predicted by the integration model than with maps predicted by either of the alternative nonintegration strategies (Figure 5G; twotailed ttest on bootstrapped r values: p < 0.001 for six out of nine participants in the integration vs. snapshot comparison; in all nine participants for the integration vs. extremadetection comparison). Overall, these analyses provide converging evidence that human decisions in an orientation discrimination task rely on temporal integration.
Auditory intensity discrimination in rats relies on temporal integration
Finally, we analyzed data from five rats performing a fixedduration auditory task where the animals had to discriminate the side with larger acoustic intensity (PardoVazquez et al., 2019). The relative intensity of the left and right acoustic signals was modulated in sensory samples of 50ms, so that the stimulus sequence provided timevarying evidence for the rewarded port. The stimulus sequence was composed of either 10 or 20 acoustic samples of 50 ms each, for a total duration of 500 or 1000 ms. We applied the same analysis pipeline as for monkey and human data. The integration model provided a much better account of rat choices than nonintegration strategies, based on psychometric curves (Figure 6A), predicted accuracy (Figure 6B), psychophysical kernel (Figure 6C), and model comparison using AIC (Figure 6D). Similar to humans and monkeys, rats tended to select the side corresponding to the total stimulus evidence and not the largest sample evidence in ‘disagree’ trials, as predicted by the integration model (Figure 6E). Finally, the integration map was largely consistent with an integration strategy (Figure 6F), and correlated more strongly with simulated maps from the integration model (unpaired ttest on bootstrapped r values: p < 0.001 for each animal and comparison against extremadetection and against snapshot model).
Discussion
We investigated the presence of temporal integration in perceptual decisions in monkeys, humans, and rats through a series of standard and innovative analyses of response patterns. In all analyses, we contrasted predictions from one integration and two nonintegration computational models of behavioral responses (Figure 1). For each nonintegration model, we considered multiple variants to explore the maximal flexibility offered by each framework to capture animal behavior. For our datasets, evidence in favor of integration was easy to achieve using standard model comparison techniques as well as comparing simulated psychometric curves and psychophysical kernels to their experimental counterparts (Figure 2). Our results are in line with previous evidence for temporal integration in perceptual decisions of humans and mice (Pinto et al., 2018; Stine et al., 2020; Waskom and Kiani, 2018). Importantly, we also put forth new analyses targeted at revealing specific signatures of temporal integration.
In some cases, we could link the failure of the nonintegration model to a fundamental limitation of the model. For example, the extremadetection model cannot explain the nonmonotonic psychophysical kernels of monkeys or the increasing psychophysical kernels in humans. This is because the decision in that mode is based on the first sample to reach a certain fixed criterion, so it will always produce a primacy effect, that is, a decreasing psychophysical kernel. Although this effect can be small, and in practice yields approximately flat kernels (Stine et al., 2020), it cannot produce increasing or nonmonotonic kernels.
Another strong limitation of nonintegration models (both the extremadetection and the snapshot model) is that accuracy is limited by the fact that decisions depend on a single sample. We found that that boundary performance (i.e., the maximum performance that a model can reach) was actually lower than subject accuracy for most human participants, de facto ruling out these nonintegration strategies for these participants. This is consistent to what was observed in a contrast discrimination DSS task where human subjects had to make judgments about image sequences spanning up to tens of seconds each (Waskom and Kiani, 2018). It clearly contrasts however with results from Stine et al., 2020 where the nonintegration strategies matched the accuracy of human subjects performing the classical randomdotmotion task. This discrepancy may be related to the different sources of noise in the two paradigms. In DSS tasks, because the sensory evidence provided by the stimulus at each moment is controlled by the experimenter, the unpredictability of human responses essentially stems from internal noise at the level of sensory processing and temporal integration (Waskom and Kiani, 2018; Drugowitsch et al., 2016). By contrast, the random dot motion task (Kiani et al., 2008), which is a nonDSS task because the experimenter does not typically specify differing amounts of motion in each time epoch within a single trial, typically elicits more variable responses due to the presence of stimulus noise. This overall increased noise level leads to a looser relationship between the stimulus condition and the behavioral responses, which can thus be accounted for by a larger spectrum of computational mechanisms. These issues have been addressed by forcing ‘pulses’ of a certain stimulus strength and/or by performing post hoc analyses to estimate signal and noise (Kiani et al., 2008; Huk and Shadlen, 2005) but these are partial solutions that DSS paradigms solve by design. This illustrates the benefits of using experimental designs where variability in stimulus information can be fully controlled and parametrized by the experimenter, as these paradigms discriminate more precisely between different models of perceptual decisions (see Cisek et al., 2009).
In at least one monkey, although quantitative metrics such as penalized loglikelihood and fits to psychometric curves clearly pointed to the integration model as the best account to behavior, the qualitative failure modes of the nonintegration strategies (especially the snapshot model) was not immediately clear. Although we tried variants for each nonintegration model, there remained a possibility that our precise implementation failed to account for monkey behavior but that other possible implementations would. Note that the extremadetection and snapshot are two of the many possible nonintegration strategies. A generic form for nonintegration strategies corresponds to a policy that implements positiondependent thresholds on the instantaneous sensory evidence. In this framework, the extremadependent model corresponds to the case with a positionindependent threshold, while the snapshot model corresponds to a null bound for one sample and infinite bounds for all other samples. To rule out these more complex strategies, we conducted additional analyses that specifically targeted core assumptions of the integration and nonintegration strategies.
First, the extremadetection model fails to account for the data because it predicts that largest evidence samples should have a disproportionate impact on choices. However, this does not occur, as monkeys and humans tend to respond according to the total evidence and not the single largeevidence sample (Figures 3C and 5E) – and see Levi et al., 2018 for a similar analysis. All nonintegration strategies share the property that on each trial the decision should only rely either on the early or the late part of the trial. We thus directly examined the assumptions of integration and nonintegration models by assessing how the evidence from the early and late parts of each stimulus sequence is combined to produce a decision. We introduced integration maps (Figure 4) to inspect such integration: isolines of the integration maps will be rectilinear if and only if early and late evidence are summed, in other words if and only if temporal integration takes place. Unequal weighting of evidence would still produce rectilinear isolines, albeit with a different angle. By contrast, a nonintegration scenario when on each trial only a single piece of evidence contributes to the decision predicts isolines that bend toward the axes. Integration maps from monkey, human, and rat subjects nicely matched the predictions of the integration models, proving that their decisions do rely on temporal integration. Note that this innovative analysis technique could be used to probe integration of evidence not only at temporal level but also between different sources of evidence. Indeed, there has been an intense debate about whether sensory information from different spatial locations or different modalities are integrated prior to reaching a decision, or whether decisions are taken separately for each source before being merged, which can be viewed as extensions to the snapshot model (Pannunzi et al., 2015; Otto and Mamassian, 2012; Lorteije et al., 2015; Hyafil and MorenoBote, 2017). Our integration analysis could provide new answers to this old debate.
Integration maps can be computed not only for choice patterns but for any type of behavioral or neural marker of cognition. We computed a neural integration map (Figure 4—figure supplement 3) by looking at the average spike activity of Lateral Intra Parietal (LIP) neurons as a function of early and late evidence, for neurons recorded while the monkeys performed the motion discrimination experiment (Yates et al., 2017). The neural integration map clearly showed rectilinear isolines, as predicted by an integration model of neural spiking. By contrast, neural implementations of the snapshot and extremadetection predicted strongly curved isolines. The activity of LIP neurons correlates with the evidence accumulated over the presentation of the stimulus in favor of either possible choices (Gold and Shadlen, 2007). This result shows that the activity of individual LIP neurons indeed reflects the temporal integration of sensory information that drives animal behavior.
We have focused in this study on paradigms where the stimulus duration is fixed by the experimenter, and subjects could only respond after stimulus extinction. Stine et al. proposed a method for distinguishing integration from nonintegration strategies but the method is based on two experimental conditions: experiments where stimulus duration is controlled by the experimenter and experiments where the stimulus is controlled by the subject (i.e., plays until the subject responds, ‘reaction time paradigms’). Our study puts forth a method to differentiate integration from nonintegration strategies that is based on only one experimental condition (i.e., the variableduration paradigm), and may therefore be applied to existing datasets.
Other studies have shown how integration and nonintegration strategies can be disentangled in free reaction time task paradigms. Specifically, different models make different predictions regarding how the total sample evidence presented before response time should vary with response time (Glickman and Usher, 2019; Zuo and Diamond, 2019). Glickman and Usher used these predictions to rule out nonintegration strategies in a counting task in humans, and Zuo and Diamond, 2019 found evidence for evidence integration to bound when rats discriminate textures using whisker touches. Under strong urgency constraints, it has been proposed that decisions depend on very limited temporal integration of the stimulus by lowpass filtering of stimulus evidence (Cisek et al., 2009; Thura et al., 2012). However, this suggestion cannot explain the fact that evidence presented early in the trial influences decisions taken later on Winkel et al., 2014; Thura et al., 2012. Of note, the absence of integration seems a more viable strategy when the duration of the stimulus is controlled externally and the benefits of integrating in terms of accuracy might not compensate for its cognitive cost. In free reaction time paradigms, waiting for a long sequence of samples and selecting its response based on a single sample does not seem a particularly efficient strategy. If the cognitive cost of integration is high, it is more beneficial to interrupt the stimulus sequence early with a rapid response. Indeed, fast or very fast (under 250 ms) perceptual decisions are very common (Uchida et al., 2006; Zariwala et al., 2013; Stanford and Salinas, 2021; PardoVazquez et al., 2019; HermosoMendizabal et al., 2020). Such rapid termination of the decision process have been attributed either to urgency signals modulating the integration of stimulus evidence (Drugowitsch et al., 2012; Stanford and Salinas, 2021) or to action initiation mechanisms that time the response after a specific time (e.g., one or two samples) following stimulus onset (HernándezNavarro et al., 2021). Here, we have shown that even in paradigms where the stimulus duration is controlled by the experimenter, mammals often integrate sensory evidence over the entire stimulus.
In conclusion, we have found strong evidence for temporal integration in perceptual tasks across species (monkeys, humans, and rats) and perceptual domain (visual motion, visual orientation, and auditory discrimination). Although the timescale of integration can be adapted to the statistics of the environment (Ossmy et al., 2013; Glaze et al., 2015; Kilpatrick et al., 2019), the principle that stimulus evidence is integrated over time appears to be a hallmark of perception. This evidence was gathered by leveraging experimentally controlled sensory evidence at each sensory sample composing a stimulus, and novel modelbased statistical analysis. We speculate that temporal integration is a ubiquitous feature of perceptual decisions due to hardwired neural integrating circuits, such as recurrent stabilizing connectivity in sensory and perceptual areas (Wang, 2008; Wimmer et al., 2015).
Methods
Monkey experiment
We present here the most relevant features of the behavioral protocol – see Yates et al., 2017 for further experimental details. Two adult rhesus monkeys performed a motion discrimination task. On each trial, a stimulus consisting of a hexagonal grid (5–7 degrees, scaled by eccentricity) of Gabor patches (0.9 cycle per degree; temporal frequency 5 Hz for Monkey P; 7 Hz for Monkey N) was presented. Monkeys were trained to report the net direction of motion in a field of drifting and flickering Gabor elements with an eye movement to one of two targets. Each trial motion stimulus consisted of seven consecutive motion pulses, each lasting 9 or 10 video samples (150 or 166 ms; pulse duration did not vary within a session), with no interruptions or gaps between the pulses. The strength and direction of each pulse ${S}_{ti}$ for trial t and sample i were set by a draw from a Gaussian rounded to the nearest integer value. The difficulty of each trial was modulated by manipulating the mean and variance of the Gaussian distribution. Monkeys were rewarded based on the empirical stimulus and not on the stimulus distribution. We analyzed a total of 112 sessions for monkey N and 60 sessions for monkey P, with a total of 72,137 and 33,416 valid trials, respectively. These sessions correspond to sessions with electrophysiological recordings reported in Yates et al., 2017 and purely behavioral sessions.
Human experiment
Nine adult subjects performed an orientation discrimination task whereby on each trial they reported in each trial whether a series of gratings were perceived to be mostly tilted clockwise or counterclockwise (Drugowitsch et al., 2016). Each DSS consisted of 5–10 gratings. Each grating was a highcontrast Gabor patch (color: blue or purple; spatial frequency = 2 cycles per degree; SD of Gaussian envelope = 1 degree) presented within a circular aperture (4 degrees) against a uniform gray background. Each grating was presented during 100 ms, and the interval between gratings was fixed to 300 ms. The angles of the gratings were sampled from a von Mises distribution centered on the reference angle (${\alpha}_{0}=45$ degrees for clockwise sequences, 135 degrees for anticlockwise sequences) and with a concentration coefficient $\kappa =0.3$. The normative evidence provided by sample i in trial t in favor of the clockwise category corresponds to how well the grating orientation ${\alpha}_{ti}$ aligns with the reference orientation, that is ${S}_{ti}=2\kappa cos({2(\alpha}_{ti}{\alpha}_{0}))$ .
Each sequence was preceded by a rectangle flashed twice during 100 ms (the interval between the flashes and between the second flash and the first grating varied between 300 and 400 ms). The participants indicated their choice with a button press after the onset of a centrally occurring dot that succeeded the rectangle mask and were made with a button press with the right hand. Failure to provide a response within 1000 ms after central dot onset was classified as invalid trial. Auditory feedback was provided 250 ms after participant response (at latest 1100 ms after end of stimulus sequence). It consisted of an ascending tone (400/800 Hz; 83/167 ms) for correct responses; descending tone (400/400 Hz; 83/167 ms) for incorrect responses; a low tone (400 Hz; 250 ms) for invalid trials.
Trials were separated by a blank interstimulus interval of 1200–1600 ms (truncated exponential distribution of mean 1333 ms). Experiments consisted of 480 trials in 10 blocks of 48. It was preceded with two blocks of initiation with 36 trials each. In the first initiation block, there was only one grating in the sequence, and it was perfectly aligned with one of the reference angles. In the second initiation block, sequences of gratings were introduced, and the difficulty was gradually increased (the distribution concentration linearly decreased from $\kappa =1.2$ to $\kappa =0.3$). Invalid trials (mean 6.9 per participant, std 9.4) were excluded from all regression analyses. The study was approved by the local ethics committee (approval 2013/5435/I from CEImParc de Salut MAR).
Rat experiment
Rat experiments were approved by the local ethics committee of the University of Barcelona (Comité d’Experimentació Animal, Barcelona, Spain, protocol number Ref 390/14). Five male LongEvans rats (350–650 g), pairhoused and kept on stable conditions of temperature (23°C) and humidity (60%) with a constant light–dark cycle (12:12 hr, experiments were conducted during the light phase). Rats had free access to food, but water was restricted to behavioral sessions. Free water during a limited period was provided on days with no experimental sessions.
Rats performed a fixedduration auditory discrimination task where they had to classify noisy stimuli based on the intensity difference between the two lateral speakers (PardoVazquez et al., 2019; HermosoMendizabal et al., 2020). An LED on the center port indicated that the rat could start the trial by poking in that center port. After this poke, rats had to hold their snouts in the central port during 300 ms (i.e., fixation). Following this period, an acoustic DSS was played. Rats had to remain in the central port during the entire presentation of the stimulus. At stimulus offset, the center LED went off and rats could then come out of the center port and head toward one of the two lateral ports. Entering the lateral port associated with the speaker that generated the larger sound intensity led to a reward of 24 µl of water (correct responses), while entering the opposite port lead to a 5s timeout accompanied with a bright light during the entire period (incorrect responses). If rats broke fixation during the prestimulus fixation period or during the stimulus presentation, the sound was interrupted, the center LED remained on, and the rat had to initiate a new trial starting by center fixation followed by a new stimulus. Fixation breaks were not included in any of the analyses. Stimulus duration was 0.5 s (10 samples) or 1 s (20 samples). Two rats performed 0.5s stimuli only (77,810 and 54,803 valid trials, respectively); one rat performed 1s stimuli only (42,474 valid trials); the remaining two rats performed a mixture of 0.5 and 1 s stimuli trials randomly interleaved (5016 trials and 65,212 valid trials, respectively, for one animal; 7374 and 38,829 trials for the other animal). In each trial k one stimulus ${S}_{K}^{X}(t)$ was played in each speaker (X = R for the Right speaker and X = L for the Left speaker). Each stimulus was an amplitudemodulated (AM) broadband noise defined by ${S}_{k}^{X}\left(t\right)=[1+sin({f}_{AM}t+\phi \left)\right]{a}_{k}^{X}\left(t\right){\xi}_{X}\left(t\right)$ where f_{AM} = 20 Hz (sensory samples lasted 50 ms), the phase delay $\phi =3\pi /2$ and ${\xi}_{X}\left(t\right)$ were broadband noise bursts. The amplitudes of each sound in each frame were ${a}_{k}^{L}\left(t\right)=\left(1+{S}_{k,f}\right)/2$ and ${a}_{k}^{R}\left(t\right)=\left(1{S}_{k,f}\right)/2$ with S_{k,f}(t) being the instantaneous evidence that was drawn independently in each frame f from a transformed Beta distribution with support [−1,1]. With this parametrization of the two sounds the sum of the two envelopes was constant in all frames a^{L}_{k}(t) + a^{R}_{k}(t) = 1. There were 7 × 5 stimulus conditions, each defined by a Beta distribution, spanning 7 mean values (−1, −0.5, −0.15, 0, 0.15, 0.5, and 1) and 5 different standard deviations (0, 0.11, 0.25, 0.57, and 0.8). In around the first half of the sessions, only sample sequences in which the total stimulus evidence matched the targeted nominal evidence were used. This effectively introduced weak correlations between samples. In the second half of the sessions, this condition was removed and samples in each stimulus were drawn independently from the corresponding Beta distribution.
Integration model
The integration model for human participants corresponds to a logistic regression model, where the probability of selecting the right choice $p\left({r}_{t}\right)$ at trial t depends on the weighted sum of the sample evidence: $p\left({r}_{t}\right)\text{}=\text{}\sigma \left({\beta}_{0}+{\mathrm{\Sigma}}_{i\in \left[\mathrm{1...}n\right]}{\beta}_{i}{S}_{ti}\right)$, where ${\beta}_{0}$ is a lateral bias, ${S}_{ti}$ is the signed sample evidence at sample i; ${\beta}_{i}$ is the sensory weight associated with the ith sample in the stimulus sequence; and $\sigma \left(x\right)=({1+{e}^{x})}^{1}$ is the logistic function. The vector ${\beta}_{i}$ ’s allowed to capture different shapes of psychophysical kernels (e.g., primacy effects, recency effects) which can emerge due to a variety of suboptimalities in the integration process (leak, attractor dynamics, sticky bounds, sensory aftereffects, etc.) (Brunton et al., 2013; Yates et al., 2017; PratOrtega et al., 2021; Bronfman et al., 2016; Keung et al., 2019; Keung et al., 2020). Technically, our statistical approach corresponds to the firstorder expansion of the Volterra series used to approximate any integration model (Neri, 2004) by a simple and fittable logistic model. In other words, because a more complete description of integration models would be computationally challenging to fit and prone to overfitting, we chose a statistical approximation in the form of the logistic model that captures the essence of any generative model that include temporal integration, that is weighted summation of stimulus evidence.
For the monkey and rat data, we included a sessiondependent modulation gain ${\gamma}_{t}$ to capture the large variations in performance in monkeys across the course of sessions (see Figure 2—figure supplement 1A):
This model corresponds to a bilinear logistic regression model which pertains to the larger family of Generalized Unrestricted Models (GUMs) (Adam and Hyafil, 2020). Parameters ($\beta ,\gamma $) were fitted using the Laplace approximation as described in Adam and Hyafil, 2020. The modulation gain was omitted when applied to human data, yielding a classical logistic regression model.
Snapshot model
In the snapshot model, decisions are based on each trial based upon a single sample. The model also includes the possibility for left and right lapses. In each trial, the attended sample is drawn from a multinomial distribution of parameters ($\pi}_{1},\text{}...{\pi}_{n},\text{}{\pi}_{L},\text{}{\pi}_{R$), where the first terms ${\pi}_{i}(1\le i\le n)$ correspond to the probability of attending sample i, and ${\pi}_{L}$ and ${\pi}_{R}$ correspond to the probability of left and right lapses, respectively. Upon selecting sample i, the probability for selecting the right choice is given by the function ${H}_{i}\left({S}_{t}\right)$. In the deterministic version of the model, ${H}_{i}$ is simply determined by the sign of the ith sample evidence: ${H}_{i}\left({S}_{t}\right)=1$ if ${S}_{ti}>0$, ${H}_{i}\left({S}_{t}\right)=0$ if ${S}_{ti}<0$, and ${H}_{i}\left({S}_{t}\right)=0.5$ if ${S}_{ti}=0$ (i.e., random guess if the sample has null evidence). We also define similar functions for lapse responses: ${H}_{R}\left(S\right)=1$ and ${H}_{L}\left(S\right)=0$, irrespective of the stimulus. In the nondeterministic version of the model, the probability ${H}_{i}\left({S}_{ti}\right)$ is determined by a logistic function of the attended sample evidence ${H}_{i}\left({S}_{t}\right)=\sigma \left({{\beta}_{i}S}_{ti}\right)$ where ${\beta}_{i}$ describes a sensitivity parameter. The deterministic case can be viewed as the limit of the nondeterministic case when all sensitivity parameters ${\beta}_{i}$ diverge to $+\infty $, that is when sensory and decision noise are negligible.
The overall probability for selecting right choice (marginalizing over the attended sample, which is a hidden variable) can be captured by a mixture model:
The mixture coefficients ${\pi}_{i}\text{}\left(i=1,...n,L,R\right)$ are constrained to be nonnegative and sum up to 1. In the nondeterministic model, the parameters also include sensitivity parameters ${\beta}_{i}$ .The model is fitted using ExpectationMaximization (Bishop, 2006). In the Expectation step, we compute the responsibility ${z}_{ti}$ , that is the posterior probability that the sample i was attended at trial t (for i = L, R, the probability that the trial corresponded to a lapse trial):
In the Maximization step, we update the value of the parameters by maximizing the Expected Complete LogLikelihood (ECLL): $Q(\pi ,\beta )={\Sigma}_{ti}{z}_{ti}logp({r}_{t};\pi ,\beta )$. Maximizing over the mixture coefficients with the unitysum constraint provides the classical update: ${\pi}_{i}={\Sigma}_{ti}{z}_{ti}/N$, where N is the total number of trials. In the nondeterministic model, maximizing the ECLL over sensitivity parameters is equivalent to fitting a logistic regression model with weighted coefficients ${z}_{ti}$, which is a convex problem. Best fitting parameters can be found using Newton–Raphson updates on the parameters:
To speed up the computations, in each M step, we only performed one Newton–Raphson update for each sensitivity parameter, rather than iterating the updates fully until convergence. The EM procedure was run until convergence, assessed by an increment in the loglikelihood $L(\pi ,\beta )$ of less than 10^{−9} after one EM iteration. The loglikelihood for a given set of parameters is given by $L(\pi ,\beta )={\Sigma}_{t}logp\left({r}_{t}\right)$. The EM iterative procedure was repeated with 10 different initializations of the parameters to avoid local minima.
Note that for monkey and rat data, since we observed large variations in performance across sessions, the model based its choices on sessiongainmodulated evidence ${\stackrel{}{\mathrm{S}}}_{ti}={\gamma}_{t}{S}_{ti}$ instead of raw evidence ${S}_{ti}$ (this had no impact for the deterministic variant since ${\stackrel{}{\mathrm{S}}}_{ti}$ and ${S}_{ti}$ always have the same sign). We fitted the model from individual subject responses either with lapses ${\pi}_{L}$ and ${\pi}_{R}$ as free parameters, or fixed to ${\pi}_{L}={\pi}_{R}=0.01$. Figures in the main manuscript correspond to the deterministic snapshot model with fixed lapses. We also studied variants of the snapshot model where decisions in each trial are based on K attended samples, that is depends on $({S}_{ti},..{S}_{t,i+K1})$ with $1\le K\le n1$ and $1\le i\le nK+1$ is the first attended sample. In the deterministic case, the choice is directly determined by the sign of the sum of the signed evidence for the attended samples. In the nondeterministic case, the evidence for the attended samples are weighted and passed through a sigmoid: ${H}_{i}\left({S}_{t}\right)=\sigma \left({\mathrm{\Sigma}}_{k\in \left[\mathrm{1...}K\right]}{\beta}_{ki}{S}_{t,i+k1}\right)$. The model with a single attended sample presented above is equivalent to this extended model when using $K=1$. At the other end, using $K=n$ corresponds to the temporal integration model (without the lateral bias).
Extremadetection model
In the extremadetection model, a choice is selected according to the first sample in the sequence whose absolute evidence value reaches a certain threshold $\theta $, that is $p\left({r}_{t}\theta \right)=\text{}H\left({m}_{ti}\right),\text{}{m}_{ti}\ge \theta ,\text{}{m}_{tj}\theta$ for all $j<i$ . Here, ${m}_{ti}$ is the sample evidence corrupted by sensory noise ${\epsilon}_{ti}$ which is distributed normally with variance ${\sigma}^{2}$ : ${m}_{ti}={S}_{ti}+{\epsilon}_{ti}$ with ${\epsilon}_{ti}\sim N(0,{\sigma}^{2})$. H is the step function. If the stimulus sequence ends and no sample has reached the threshold, then the decision is taken at chance. As described in Waskom and Kiani, 2018, the probability for a rightward choice at trial t can be expressed as:
where $\Phi $ is the cumulative normal distribution. We also included the possibility for left and right lapses with probability ${\pi}_{L}$ and ${\pi}_{R}$ . Following Stine et al., 2020, we explored an alternative default rule called ‘last sample’ rule: if the stimulus extinguishes and the threshold has not been reached, then the decision is based on the (noisy) last sample rather than simply by chance. This changes the equation describing the probability for rightward choices to:
We also explored a variant of the model where the threshold changes on every sample, that is the equations above are changed by substituting $\theta $ with ${\theta}_{i}$ . As for the snapshot model, we used the sessiongainmodulated evidence ${\stackrel{}{\mathrm{S}}}_{ti}$ instead of raw evidence ${S}_{ti}$ for fitting the model to monkey and rat data. The four parameters of the model $(\theta ,\sigma ,{\pi}_{L},{\pi}_{R})$ were estimated from each subject data by maximizing the loglikelihood with interiorpoint algorithm (function fmincon in Matlab) and 10 different initializations of the parameters. (In the varyingthreshold variant, there are n + 3 parameters which are estimated similarly.)
Model validation and model comparison
Psychophysical kernels were obtained from subject and simulated data by running a logistic regression model: $p\left({r}_{t}\right)=\sigma ({\beta}_{0}+{\Sigma}_{i}{\beta}_{i}{S}_{ti})$. Standard errors of the weights ${\beta}_{i}$ were obtained from the Laplace approximation. For psychometric curves, we first defined the weighted stimulus evidence T_{t} at trial t as the sessionmodulated weighted sum of signed sample evidence; with the weights obtained from the logistic regression model above ${{T}_{t}={\gamma}_{t}\Sigma}_{i}{\beta}_{i}{S}_{ti}$ . We then divided the total stimulus evidence into 50 quantiles (10 for human subjects) and computed the psychometric curve as the proportion of rightward choices for each quantile.
The boundary performance for the snapshot and extremadetection models corresponds to the best choice accuracy out of all the parameterizations for each model. In the snapshot model, the boundary performance corresponds to the deterministic version with nolapse, where the attended sample is always the sample ${i}^{*}$ whose sign better predicts the stimulus category over all animal trials, that is ${\pi}_{{i}^{*}}=1$ and ${\pi}_{i}=0$ if ${i\ne i}^{*}$ . For the extremadetection model, the boundary performance corresponds to the lapsefree model with no sensory noise ($\sigma =0$) and a certain value for threshold $\theta $ that is identified for each subject by simple parameter search.
Finally, model selection was performed using the AIC $AIC=2p2{L}_{ML}$ , where p is the number of model parameters and ${L}_{ML}$ is the likelihood evaluated at maximum likelihood parameters.
Analysis of majoritydriven choices
We selected for each animal the subset of trials corresponding to when the largest evidence sample was at odds with the total stimulus evidence (disagree trials), that is where $sign({S}_{tj},{S}_{tj}\ge {S}_{ti}\vee i)\ne sign\left({\Sigma}_{i}{S}_{ti}\right)$. For this subset of trials, we computed the proportion of animal choices that were aligned with the overall stimulus evidence. We repeated the analysis for simulated data from the integration and extremadetection models. We computed error bars following a parametric bootstrap procedure: for each bootstrap, we simulated the model (integration/extremadetection) using parameters sampled from their posterior distribution (based on the Laplace approximation). We then applied the analysis on disagree trials for simulated data, and used these bootstrap values to define confidence intervals.
Subjective weighting analysis
In order to estimate the impact of each sample on the animal choice as a function of sample evidence, we built and estimated the following statistical model:
As can be seen, this model is equivalent to the temporal integration model under the assumption that f is a linear function. Rather, here we wanted to estimate the function f (as well as the session gain ${\gamma}_{t}$ , lateral bias ${\beta}_{0}$, and sensory weight ${\beta}_{i}$). Including the session gain was necessary for estimating f accurately from the monkey and rat behavioral data, since the distribution of pulse strength ${S}_{ti}$ was varied across sessions and could otherwise induce a confound. We assumed that f is an odd function, that is $f\left({S}_{ti}\right)=f\left({S}_{ti}\right)$. This equation takes the form of a GUM and was fitted using the Laplace approximation method as described in Adam and Hyafil, 2020. In the monkey experiment, sample evidence could take only a finite number of values, so f was simply estimated over these values. In the human experiment, sample evidence could take continuous values. In this case, we defined a Gaussian Process prior over f with squared exponential kernel with length scale 0.1 and variance 1.
Integration of early and late evidence
We designed a new analysis tool to characterize the statistical mapping from the multidimensional stimulus space ${\mathit{S}}_{t}=({S}_{t1},...{S}_{tn})\in {\mathfrak{R}}^{n}$ onto binary choices ${r}_{t}\in \left[\mathrm{0,1}\right]$. We first collapsed the stimulus sequence ${\mathit{S}}_{t}$ onto the twodimensional space defined by early evidence ${\mathit{E}}_{t}$ and late evidence ${\mathit{L}}_{t}$ defined by ${{E}_{t}={\gamma}_{t}\Sigma}_{1\le i\le [n/2]}{\beta}_{i}{S}_{ti}$ and ${{L}_{t}={\gamma}_{t}\Sigma}_{[n/2]+1\le i\le n}{\beta}_{i}{S}_{ti}$ , where the weights ${\beta}_{i}$ and session gains ${\gamma}_{t}$ correspond to parameters estimated from the temporal integration model (session gains were omitted for human participants). Next we plotted the integration map which represents the probability for rightward choices as a function of $({E}_{t},{L}_{t})$. The map was obtained by smoothing data points with a twodimensional Gaussian kernel. More specifically, for each pair value (E,L), we selected the trials whose early and late evidence values ${E}_{t}$ and ${L}_{t}$ fell within a certain distance to (E,L), that is ${d}_{t}=dist\left(\left(E,L\right)\left({E}_{t},{L}_{t}\right)\right)<\text{}2$. We then computed the proportion of rightward choices for the selected trials, with a weight for each trial depending on the distance to the pair value ${w}_{t}=N\left(\right({E}_{t},{L}_{t});(E,L),0.{1}^{2}I)$. Because the space (E,L) was not sampled uniformly during the experiment, we represent the density of trials by brightness. For each subject, we obtained integration maps both from subject data and from model simulations. For each model, we computed the Pearson correlation between the maps obtained from the corresponding simulation and from the subject data. We tested the significance of correlation measures between models by using a bootstrapping procedure: we calculated the correlation measure r from 100 bootstraps for each model and participant, and then performed an unpaired ttest between bootstrapped r.
Next, we analyzed the conditional psychometric curves, that is the psychometric curves for the early evidence conditioned on the value of late evidence, which correspond to vertical cuts in the integration map. To do so, we first binned late evidence ${L}_{t}$ by bins of width 0.5. Conditional psychometric curve represents the probability of rightward choices as a function of early evidence ${E}_{t}$, separately for each late evidence bin. For each late evidence bin, we also estimated the corresponding bias $\beta $, left lapse ${\pi}_{L}$ and right lapse ${\pi}_{R}$ by fitting the following function on the corresponding subset of trials:
Analysis of LIP neuron activity
We analyzed the activity of 82 LIP neurons recorded over 43 sessions of the motion discrimination tasks (Yates et al., 2017). We applied the following procedure to extract the integration map for LIP neurons. For each neuron n, we computed the spike count ${{s}_{t}}^{\left(n\right)}$ in a window of 500 ms width following each stimulus offset, which is where LIP neurons were found to have maximal selectivity to motion evidence from the entire pulse sequence (Yates et al., 2017). We then applied a Poisson GLM ${{E(s}_{t}}^{\left(n\right)})=exp({{w}_{0}}^{\left(n\right)}+{\Sigma}_{i}{{w}_{i}}^{\left(n\right)}{S}_{ti})$ for each neuron n to extract the impact of each sample i on the individual neural spike count ${{w}_{i}}^{\left(n\right)}$ . For each trial t, we used these weights to compute the neuronweighted early and late evidence defined by and ${{E}_{t}}^{\left(n\right)}={\Sigma}_{1\le i\le 3}{{w}_{i}}^{\left(n\right)}{S}_{ti}$ and ${{L}_{t}}^{\left(n\right)}={\Sigma}_{4\le i\le 7}{{w}_{i}}^{\left(n\right)}{S}_{ti}$ . Note that this weighting converts the evidence onto the space defined by the preferred direction of the neuron, such that positive evidence signals evidence toward the preferred direction and negative evidence signals evidence toward the antipreferred direction. We then merged the vectors for normalized spike counts ${\stackrel{}{s}}_{t}^{\left(n\right)}={{s}_{t}}^{\left(n\right)}/exp\left({{w}_{0}}^{\left(n\right)}\right)$, early evidence ${{E}_{t}}^{\left(n\right)}$ and late evidence ${{L}_{t}}^{\left(n\right)}$ across all neurons. The normalized spike counts were binned by values of early and late evidence (bin width: 0.02), and the average over each bin was computed after convolving with a twodimensional Gaussian kernel of width 0.1. The neural integration map represents the average normalized activity per bin.
Simulations of spiking data for the integration and nonintegration models were proceeded as follows. First, the neural integration model corresponds to linear summing with neuronspecific weights which are then passed through an exponential nonlinearity; the spike counts for each trial are generated using a Poisson distribution whose rate is equal to the nonlinear output (Figure 4—figure supplement 3a, top). This corresponds exactly to the generative process of the Poisson GLM described above. For the extremadetection model (Figure 4—figure supplement 3a, middle), we hypothesized that LIP activity would only be driven by the sample that reaches the threshold (and dictates the animal response). To this end, we first simulated the behavioral extremadetection model for all trials, using parameters $(\theta ,\sigma ,{\pi}_{L},{\pi}_{R})$ fitted from the corresponding animal, to identify which sample i reaches the subjectspecific threshold. We then assumed that the spiking activity of the neuron would follow the stimulus value at sample i ${S}_{ti}$ (signed by the preferred direction of the neuron ${p}^{\left(n\right)}$ through):
Again the spike count were generated from a Poisson distribution with rate ${{{E}_{ED}(s}_{t}}^{\left(n\right)})$.
Finally, for the snapshot model (Figure 4—figure supplement 3a, bottom), we assumed that the neuron activity would merely reflect the sensory value of the only sample it would attend. We assumed that the probability mass function to attend each of the seven samples would be neuron specific, so we used the normalized weights of the Poisson GLM for that specific neuron as defining such probability (weights were signed by the neuron preferred direction so that the vast majority of weights were positive; negative weights were ignored). For each trial, we thus randomly sampled the attended sample i using this probability mass function and then simulated the spike count ${{s}_{t}}^{\left(n\right)}$ from a Poisson distribution with rate
We simulated spiking activity for each neuron and for each integration and nonintegration model, and then used simulated data to compute neural integration maps exactly as described above for the actual LIP neuron activity.
Data availability
All experimental data (behavioral and neural data in monkeys, behavioral data in rats and humans) and code to run the analysis are publicly available at https://github.com/ahyafil/TemporalIntegration (copy archieved at Hyafil, 2023).
References

Decisions reduce sensitivity to subsequent informationProceedings. Biological Sciences 282:20150228.https://doi.org/10.1098/rspb.2015.0228

Nonmonotonic temporalweighting indicates a dynamically modulated evidenceintegration mechanismPLOS Computational Biology 12:e1004667.https://doi.org/10.1371/journal.pcbi.1004667

Decisions in changing conditions: the urgencygating modelThe Journal of Neuroscience 29:11560–11571.https://doi.org/10.1523/JNEUROSCI.184409.2009

The cost of accumulating evidence in perceptual decision makingThe Journal of Neuroscience 32:3612–3628.https://doi.org/10.1523/JNEUROSCI.401011.2012

The neural basis of decision makingAnnual Review of Neuroscience 30:535–574.https://doi.org/10.1146/annurev.neuro.29.051605.113038

Response outcomes gate the impact of expectations on perceptual decisionsNature Communications 11:1057.https://doi.org/10.1038/s4146702014824w

A distinct mechanism of temporal integration for motion through depthThe Journal of Neuroscience 35:10212–10216.https://doi.org/10.1523/JNEUROSCI.003215.2015

Regulation of evidence accumulation by pupillinked arousal processesNature Human Behaviour 3:636–645.https://doi.org/10.1038/s4156201905514

Optimal models of decisionmaking in dynamic environmentsCurrent Opinion in Neurobiology 58:54–60.https://doi.org/10.1016/j.conb.2019.06.006

Interpreting temporal Dynamics during sensory decisionmakingCurrent Opinion in Physiology 16:27–32.https://doi.org/10.1016/j.cophys.2020.04.006

Motor contributions to the temporal precision of auditory attentionNature Communications 5:5255.https://doi.org/10.1038/ncomms6255

Estimation of nonlinear psychophysical kernelsJournal of Vision 4:82–91.https://doi.org/10.1167/4.2.2

Psychophysically measured task strategy for disparity discrimination is reflected in V2 neuronsNature Neuroscience 10:1608–1614.https://doi.org/10.1038/nn1991

Noise and correlations in parallel perceptual decision makingCurrent Biology 22:1391–1396.https://doi.org/10.1016/j.cub.2012.05.031

Deconstructing multisensory enhancement in detectionJournal of Neurophysiology 113:1800–1818.https://doi.org/10.1152/jn.00341.2014

The mechanistic Foundation of Weber’s lawNature Neuroscience 22:1493–1502.https://doi.org/10.1038/s4159301904397

An accumulationofevidence task using visual pulses for mice navigating in virtual realityFrontiers in Behavioral Neuroscience 12:36.https://doi.org/10.3389/fnbeh.2018.00036

Flexible categorization in perceptual decision makingNature Communications 12:1283.https://doi.org/10.1038/s4146702121501z

Multisensory decisionmaking in rats and humansThe Journal of Neuroscience 32:3726–3735.https://doi.org/10.1523/JNEUROSCI.499811.2012

Urgent decision making: resolving visuomotor interactions at high temporal resolutionAnnual Review of Vision Science 7:323–348.https://doi.org/10.1146/annurevvision100419103842

Decision making by urgency gating: theory and experimental supportJournal of Neurophysiology 108:2912–2930.https://doi.org/10.1152/jn.01071.2011

Seeing at a glance, smelling in a whiff: rapid forms of perceptual decision makingNature Reviews. Neuroscience 7:485–491.https://doi.org/10.1038/nrn1933

Early evidence affects later decisions: why evidence accumulation is required to explain response time dataPsychonomic Bulletin & Review 21:777–784.https://doi.org/10.3758/s1342301305518

Functional dissection of signal and noise in MT and lip during decisionmakingNature Neuroscience 20:1285–1292.https://doi.org/10.1038/nn.4611
Decision letter

Emilio SalinasReviewing Editor; Wake Forest School of Medicine, United States

Michael J FrankSenior Editor; Brown University, United States

Emilio SalinasReviewer; Wake Forest School of Medicine, United States
Our editorial process produces two outputs: i) public reviews designed to be posted alongside the preprint for the benefit of readers; ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.
Decision letter after peer review:
Thank you for submitting your article "Temporal integration is a robust feature of perceptual decisions" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, including Emilio Salinas as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Michael Frank as the Senior Editor.
The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.
Essential revisions:
1) Reviewer 2 pointed out that the model used by the authors is one of a broader family of possible models. Consideration of this broader family, at least in the form of some discussion, would strengthen the results, or would at least situate them more clearly in the theoretical landscape. See the full recommendation of Reviewer 2 below.
Reviewer #2 (Recommendations for the authors):
I have one main suggestion. Otherwise, I think this is a great study.
A very general family of models of integration of a variable "x" in response to a noisy stimulus "s" would be:
tau. dx/dt = s(t). F(x,t) + G(x,t) where G(x,t) would represent any dynamics in the absence of stimulus and F(x,t) would weight evidence according to time and current level of integrated evidence. It is unclear to this reader if the results have placed any restrictions on the potential forms of G(x,t) – the suggestion is that it should be zero, but this has not been tested I think – nor on whether the function "F()" which they show depends on time, "t" could also depend on the level of integrated evidence, "x". Since the model only includes an F(t) but then contrasts with very different models, there is an implicit suggestion or assumption that F(x,t) = F(t) and G(x,t) = 0, but this assumption is not tested. It would be nice to see an explicit test or a clear statement on what limits or lack thereof must be present in this integrationlike process. For example, a comparison with the urgencygating model of Cisek (J Neurosci 2009) would be valuable, along with a discussion of whether the authors expect different results in free response paradigms.
Supp Figure 2: This needs a bit more explaining. Legend in A is different from the caption. In B, with ΔAIC, shouldn't the AIC be based on the likelihood of the datagiven model? So why is there a black curve for "data"? In fact, how are their 3 curves at all as a ΔAIC is a difference between two values? Confusing. Perhaps this is just "AIC" and the black curve is the integration model, not the experimental data?
https://doi.org/10.7554/eLife.84045.sa1Author response
Essential revisions:
Reviewer #2 (Recommendations for the authors):
I have one main suggestion. Otherwise, I think this is a great study.
A very general family of models of integration of a variable "x" in response to a noisy stimulus "s" would be:
tau. dx/dt = s(t). F(x,t) + G(x,t) where G(x,t) would represent any dynamics in the absence of stimulus and F(x,t) would weight evidence according to time and current level of integrated evidence. It is unclear to this reader if the results have placed any restrictions on the potential forms of G(x,t) – the suggestion is that it should be zero, but this has not been tested I think – nor on whether the function "F()" which they show depends on time, "t" could also depend on the level of integrated evidence, "x". Since the model only includes an F(t) but then contrasts with very different models, there is an implicit suggestion or assumption that F(x,t) = F(t) and G(x,t) = 0, but this assumption is not tested. It would be nice to see an explicit test or a clear statement on what limits or lack thereof must be present in this integrationlike process. For example, a comparison with the urgencygating model of Cisek (J Neurosci 2009) would be valuable, along with a discussion of whether the authors expect different results in free response paradigms.
We thank the reviewer for pointing to this very important point. While the two nonintegration models were relatively easy to formalize into generic generative models that can later be fitted to experimental data, the same does not hold for the integration model. As rightfully pointed out, the most general generative model of integration should include dependence of the drift term on both time and x, both for the stimulussensitive part F(x,t) and G(x,t). This wider class of models could thus accommodate timevarying sticky bounds theta(t) (i.e. F(theta(t),t)=G(theta(t),t)=0), leak or attractive dynamics (G(x,t) = λ*x) and more. Now, fitting such an immense class of models to experimental data becomes very challenging computationally and prone to overfitting (observed values consist of 50050 000 binary observations). For this reason we decided to restrict to a much highlevel description of the computation that does not characterize the functions G() or F(), and that should not be interpreted literally as a generative model of behavior but rather as a (firstorder) statistical description for any of the generative integration model corresponding to the larger integration family. Thus, to connect with the reviewer’s public review, we do not claim that the nonmonotonic weights of the logistic model correspond genuinely to a nonmonotonic sensitivity F(t), but are rather the statistical linear description of more complex nonlinear phenomena. For example, nonmonotonic kernels (both bellshaped and Ushaped) have been described in previous studies and linked theoretically to either attractor dynamics (i.e. double well potential, G(x,t) = a.x^3 + b.x in PratOrtega et al., 2021; or timedependent dynamics in Bronfman, Brezis, and Usher 2016) or divisive normalization (Keung et al. 2019 and 2020). Technically, our statistical approach corresponds to the firstorder expansion of the Volterra series used to approximate any integration model (Neri, 2004) by a simple and fittable logistic model.
Importantly our objective in this paper was not to compare and identify these complex mechanisms at play during evidence integration, but rather to compare nonintegration models against any integration strategy. We believe that our firstorder statistical modelling of integration strategies is suited for this purpose.
References
Keung, W., Hagen, T.A. & Wilson, R.C. A divisive model of evidence accumulation explains uneven weighting of evidence over time. Nat Commun 11, 2160 (2020). https://doi.org/10.1038/s41467020156300
Keung, Waitsang, Todd A. Hagen, and Robert C. Wilson. 2019. “Regulation of Evidence Accumulation by PupilLinked Arousal Processes.” Nature Human Behaviour 3 (6): 636–45.
Neri, P. (2004). Estimation of nonlinear psychophysical kernels. Journal of Vision, 4(2), 2. https://doi.org/10.1167/4.2.2
PratOrtega, G., Wimmer, K., Roxin, A. et al. Flexible categorization in perceptual decision making. Nat Commun 12, 1283 (2021). https://doi.org/10.1038/s4146702121501z
We have added the following clarifications in the manuscript: “Departures from optimality in the accumulation process such as accumulation leak, categorization dynamics, sensory adaptation or sticky boundaries may however yield unequal weighting of the different samples (Yates et al. 2017; Brunton, Botvinick, and Brody 2013; PratOrtega et al. 2021; Bronfman, Brezis, and Usher 2016; Keung, Hagen, and Wilson 2020, 2019). To accommodate for these, we allowed the model to take any arbitrary weighting of the samples: p(r = A) = Φ(β0 + ΣiβiSi) (see Methods for details). The mapping from final accumulated evidence to choice was probabilistic, to account for the effects of noise from different sources in the decisionmaking process (Drugowitsch et al. 2016). Thus the model represented an approximate statistical description for any generative model relying on temporal integration of the stimulus evidence.” (Results, p7)
“The vector βi’s allowed to capture different shapes of psychophysical kernels (e.g. primacy effects, recency effects) which can emerge due to a variety of suboptimalities in the integration process (leak, attractor dynamics, sticky bounds, sensory aftereffects, etc.) (Brunton, Botvinick, and Brody 2013; Yates et al. 2017; PratOrtega et al. 2021; Bronfman, Brezis, and Usher 2016; Keung, Hagen, and Wilson 2019, 2020). Technically, our statistical approach corresponds to the firstorder expansion of the Volterra series used to approximate any integration model (Neri 2004) by a simple and fittable logistic model. In other words, because a more complete description of integration models would be computationally challenging to fit and prone to overfitting, we chose a statistical approximation in the form of the logistic model that captures the essence of any generative model that include temporal integration, i.e. weighted summation of stimulus evidence.” (Methods, p22).
Finally, we have expanded the discussion about freeresponse paradigms in the Discussion to include an explicit comparison with the urgencygating model of Cisek et al. 2009 (in the previous version, we referred to Thura et al. 2012 which is a more thorough theoretical development of the urgencygating model):
“Other studies have shown how integration and nonintegration strategies can be disentangled in free reactiontime task paradigms. Specifically, different models make different predictions regarding how the total sample evidence presented before response time should vary with response time (Glickman and Usher 2019; Zuo and Diamond 2019). Glickman and Usher used these predictions to rule out nonintegration strategies in a counting task in humans, and Zuo and Diamond found evidence for evidence integration to bound when rats discriminate textures using whisker touches (Zuo and Diamond 2019). Under strong urgency constraints, it has been proposed that decisions depend on very limited temporal integration of the stimulus by lowpass filtering of stimulus evidence (Cisek et al. 2009; Thura et al. 2012). However this suggestion cannot explain the fact that evidence presented early in the trial influences decisions taken later on (Winkel et al. 2014). Of note, the absence of integration seems a more viable strategy when the duration of the stimulus is controlled externally and the benefits of integrating in terms of accuracy might not compensate for its cognitive cost. In freereaction time paradigms, waiting for a long sequence of samples and selecting its response based on a single sample does not seem a particularly efficient strategy. If the cognitive cost of integration is high, it is more beneficial to interrupt the stimulus sequence early with a rapid response. Indeed, fast or very fast (under 250 ms) perceptual decisions are very common (Uchida, Kepecs, and Mainen 2006; Zariwala et al. 2013; Stanford and Salinas 2021). Such rapid termination of the decision process have can been attributed either to urgency signals modulating the integration of stimulus evidence (Drugowitsch et al. 2012; Stanford and Salinas 2021) or to action initiation mechanisms that time the response after a specific time (e.g. one or two samples) following stimulus onset (HernándezNavarro et al. 2021). Here, we have shown that even in paradigms where the stimulus duration is controlled by the experimenter, mammals often integrate sensory evidence over the entire stimulus.”
Supp Figure 2: This needs a bit more explaining. Legend in A is different from the caption.
Thank you for spotting multiple problems with this figure caption. Caption for panel A has been corrected and is now in agreement with the legend. (This figure is now labeled ‘Figure 2 figure supplement 2).
In B, with ΔAIC, shouldn't the AIC be based on the likelihood of the datagiven model? So why is there a black curve for "data"? In fact, how are their 3 curves at all as a ΔAIC is a difference between two values? Confusing. Perhaps this is just "AIC" and the black curve is the integration model, not the experimental data?
There are actually 2x2 = 4 different curves for each of the different variants of the snapshot model (deterministic vs probabilistic, and free vs fixed lapse parameters), but two curves (probabilisticfree and probabilistic free) are virtually identical. Each curve corresponds to ΔAIC that measures the difference in AIC for this specific variant of the snapshot model w.r.t the integration model (i.e. the integration model always corresponds to ΔAIC = 0). So in essence the panel compares 4 x 6 = 24 variants of the snapshot model with the integration model, and the integration model always wins (for both animals). We have clarified this point in the caption:
“AIC difference between each of the four variants of the snapshot model and the integration model. Legend as in A (full/dashed lines for fixed/free lapse parameters; black/blue curves for probabilistic/deterministic variants). Note that the probabilistic variant with either fixed or free lapses provide virtually indistinguishable values. Positive values indicate that the snapshot model provides a worse fit than the integration model.”
https://doi.org/10.7554/eLife.84045.sa2Article and author information
Author details
Funding
Agencia Estatal de Investigación (RYC201723231)
 Alexandre Hyafil
Ministerio de Economía y Competitividad (SAF201570324R)
 Jaime de la Rocha
European Research Council (ERC2015CoG683209)
 Jaime de la Rocha
National Institutes of Health (R01EY017366)
 Alexander C Huk
 Jonathan W Pillow
National Institutes of Health (NS104899)
 Jonathan W Pillow
Simons Collaboration for the Global Brain (SCGB AWD543027)
 Jonathan W Pillow
The funders had no role in study design, data collection, and interpretation, or the decision to submit the work for publication.
Acknowledgements
The authors thank Jake Yates for sharing information regarding the monkey experimental data. The authors are supported by the Spanish State Research Agency (RYC201723231 to AH), Spanish Ministry of Economy and Competitiveness together with the European Regional Development Fund grant SAF201570324R (to JR), European Research Council grant ERC2015CoG683209 (to JR), NIH grant R01EY017366 (to ACH and JWP), NIH BRAIN Initiative grant NS104899 (to JWP), and the Simons Collaboration on the Global Brain (SCGB AWD543027, JWP).
Ethics
Informed consent was obtained from all participants. The experiment with human participants was approved by the UPF ethics committee (approval 654 2013/5435/I from CEImParc de Salut MAR).
Rat experiments were approved by the local ethics committee of the University of Barcelona 658 (Comitéd'Experimentació Animal, Barcelona, Spain, protocol number Ref 390/14). Monkey experiment: All experimental protocols were approved by The University of Texas Institutional Animal Care and Use Committee (AUP201200085, AUP201500068) and in accordance with National Institutes of Health standards for care and use of laboratory animals.
Senior Editor
 Michael J Frank, Brown University, United States
Reviewing Editor
 Emilio Salinas, Wake Forest School of Medicine, United States
Reviewer
 Emilio Salinas, Wake Forest School of Medicine, United States
Publication history
 Received: October 8, 2022
 Preprint posted: October 26, 2022 (view preprint)
 Accepted: May 3, 2023
 Accepted Manuscript published: May 4, 2023 (version 1)
 Version of Record published: May 23, 2023 (version 2)
Copyright
This is an openaccess article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.
Metrics

 576
 Page views

 99
 Downloads

 0
 Citations
Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading

 Neuroscience
One signature of the human brain is its ability to derive knowledge from language inputs, in addition to nonlinguistic sensory channels such as vision and touch. How does human language experience modulate the mechanism by which semantic knowledge is stored in the human brain? We investigated this question using a unique human model with varying amounts and qualities of early language exposure: early deaf adults who were born to hearing parents and had reduced early exposure and delayed acquisition of any natural human language (speech or sign), with early deaf adults who acquired sign language from birth as the control group that matches on nonlinguistic sensory experiences. Neural responses in a semantic judgment task with 90 written words that were familiar to both groups were measured using fMRI. The deaf group with reduced early language exposure, compared with the deaf control group, showed reduced semantic sensitivity, in both multivariate pattern (semantic structure encoding) and univariate (abstractness effect) analyses, in the left dorsal anterior temporal lobe (dATL). These results provide positive, causal evidence that language experience drives the neural semantic representation in the dATL, highlighting the roles of language in forming human neural semantic structures beyond nonverbal sensory experiences.

 Neuroscience
Across phyla, males often produce speciesspecific vocalizations to attract females. Although understanding the neural mechanisms underlying behavior has been challenging in vertebrates, we previously identified two anatomically distinct central pattern generators (CPGs) that drive the fast and slow clicks of male Xenopus laevis, using an ex vivo preparation that produces fictive vocalizations. Here, we extended this approach to four additional species, X. amieti, X. cliivi, X. petersii, and X. tropicalis, by developing ex vivo brain preparation from which fictive vocalizations are elicited in response to a chemical or electrical stimulus. We found that even though the courtship calls are speciesspecific, the CPGs used to generate clicks are conserved across species. The fast CPGs, which critically rely on reciprocal connections between the parabrachial nucleus and the nucleus ambiguus, are conserved among fastclick species, and slow CPGs are shared among slowclick species. In addition, our results suggest that testosterone plays a role in organizing fast CPGs in fastclick species, but not in slowclick species. Moreover, fast CPGs are not inherited by all species but monopolized by fastclick species. The results suggest that speciesspecific calls of the genus Xenopus have evolved by utilizing conserved slow and/or fast CPGs inherited by each species.