Aversive outcomes impact human olfactory discrimination learning and generalization

Abstract

Learning associations between sensory stimuli and outcomes, and generalizing these associations to novel stimuli, is a fundamental feature of adaptive behavior. Given a noisy olfactory world, stimulus generalization holds unique relevance for the olfactory system. Recent studies suggest that aversive outcomes induce wider generalization curves by modulating discrimination thresholds, but evidence for similar processes in olfaction does not exist. Here, we use a novel olfactory discrimination learning paradigm to address the question of how outcome valence impacts associative learning and generalization in humans. Subjects underwent discrimination learning, where they learned to associate odor mixtures with either aversive (shock) or neutral (air puff) outcomes. We find better olfactory learning for odors associated with aversive compared to neutral outcomes. We further show that generalization gradients are also modulated by outcome valence, with the shock group exhibiting a steeper gradient. Computational modeling revealed that differences in generalization are driven by a narrower excitatory gradient in the shock group, indicating more discriminatory responses. These findings provide novel evidence that olfactory learning and generalization are strongly affected by the valence of outcomes. This adaptive mechanism allows for behavioral flexibility in novel situations with related stimuli and with outcomes of different valences. Because odor stimuli differ considerably from one encounter to the next, adaptive generalization may be especially important in the olfactory system.

INTRODUCTION

The ability to learn predictive relationships between sensory stimuli and biologically significant outcomes is important for adaptive behavior (Fanselow & Poulos, 2005; Mackintosh, 1983). Through Pavlovian conditioning, humans and other animals form associations between an initially neutral sensory stimulus (conditioned stimulus, CS) and a salient outcome (unconditioned stimulus, US). Once the CS has become a predictor of the US (Maren, 2001; Phelps & LeDoux, 2005), these associations can be used to adapt behavior when re-encountering the stimulus. At the same time, due to stimulus variations that naturally occur in the environment, combined with noise in our sensory systems, we rarely encounter the same sensory stimulus twice. Rather, we are more commonly faced with stimuli that resemble, but are not identical to, the originally learned CS. Mechanisms of stimulus generalization promote stimulus constancy, such that the brain can map minor differences in associative or perceptual features onto the originally learned stimulus (Bouton, 2006; Dymond, Dunsmoor, Vervliet, Roche, & Hermans, 2015; Ghirlanda & Enquist, 2003; Guttman & Kalish, 1956; McLaren & Mackintosh, 2002; Shepard, 1987; Spence, 1937).

Recent work in humans has examined generalization based on the perceptual similarity of visual and auditory stimuli (Dunsmoor, Kroes, Braren, & Phelps, 2017; Dunsmoor & LaBar, 2013; Dunsmoor & Murphy, 2014; Dunsmoor & Paz, 2015; Holt et al., 2014; Kahnt, Park, Burke, & Tobler, 2012; Kahnt & Tobler, 2016; Lissek et al., 2008; Norrholm et al., 2014; Schechtman, Laufer, & Paz, 2010). In these experiments, behavioral and neural responses to test stimuli that vary along a single stimulus dimension (e.g., auditory frequencies or visual angular orientations) are probed after discrimination training on two stimuli in this dimension (i.e., CS+ vs CS−). A key feature of stimulus generalization is the peak shift (Derenne, 2010; Hanson, 1959; Purtle, 1973; Wisniewski, Church, & Mercado, 2009). While generalization has been examined following single-stimulus training, the majority of studies investigating stimulus generalization in humans have focused on intra-dimensional discrimination paradigms (Dunsmoor & LaBar, 2013; Ghirlanda & Enquist, 2003; Hanson, 1959; Kahnt et al., 2012; Kahnt & Tobler, 2016), that is, experiments involving discrimination training between two stimuli that vary along the dimension in which generalization is tested. Such training procedures reliably result in peak shifts, i.e., the observation that responding to test stimuli during a generalization test is not maximal for the CS+, but instead to stimuli that are dissimilar to the CS+, in the direction away from the CS− (Hearst, 1968; Spence, 1937). Peak shifts are thought to result from the difference between excitatory and inhibitory gradients surrounding the CS+ and CS−, respectively (Ghirlanda & Enquist, 2003; Pearce, Esber, George, & Haselgrove, 2008; Spence, 1937).

Similar mechanisms may come into play whenever an odor wafts past the nose. The sense of smell plays an important role in guiding biologically significant behaviors such as food search, mate selection, and predator avoidance. The odors we naturally encounter are often complex mixtures of dozens to hundreds of odorant molecules (Counet, Callemien, Ouwerx, & Collin, 2002; Goff & Klee, 2006; Ohloff, 1994; Rossiter, 1996; Stevenson & Wilson, 2007), leaving considerable room for variations in stimulus configurations. Thus, the perceptual profile of odors emitted from a single source may differ markedly over time (Cain, 1992; Polak, 1973). This variation, and the fact that olfaction is an especially noisy sense, renders stimulus generalization of great relevance for the olfactory domain. Olfactory generalization has been studied in animals (Bhagavan & Smith, 1997; Bos, Roussel, Giurfa, & d’Ettorre, 2014; Chen, Barnes, & Wilson, 2011; Cleland, Morse, Yue, & Linster, 2002; Cleland & Narla, 2003; Firestein, Picco, & Menini, 1993; Wright, Kottcamp, & Thomson, 2008; Wright & Smith, 2004), but little is known about olfactory generalization in humans.

Previous work has shown that outcome valence can have a prominent effect on associative learning and generalization (Dunsmoor, Kragel, Martin, & LaBar, 2014; Dunsmoor et al., 2017; Dunsmoor & Murphy, 2014; Schechtman et al., 2010; Staddon, 1983). Aversive outcomes typically widen generalization in auditory and visual domains, expanding the range of stimuli evoking behavioral responses similar to the CS+ (Dunsmoor et al., 2014; Dunsmoor & Murphy, 2014; Schechtman et al., 2010), while having little to no effect on learning (Baldi, Lorenzini, & Bucherelli, 2004; Dunsmoor et al., 2017; Laxmi, Stork, & Pape, 2003). These results suggest that aversive outcomes relax perceptual discrimination thresholds, with the effect of impairing stimulus discrimination, while favoring stimulus generalization (Dunsmoor & LaBar, 2013; Laufer & Paz, 2012; Resnik, Sobel, & Paz, 2011; Shalev, Paz, & Avidan, 2018). Whether these profiles in the auditory and visual domains apply to the olfactory domain is not well understood. Indeed, we have previously shown that associative learning with aversive outcomes improves – rather than impedes – olfactory discrimination (Li, Howard, Parrish, & Gottfried, 2008).

Here, we devised a novel olfactory generalization paradigm to test the effect of outcome valence on stimulus generalization. We designed binary odor mixtures to systematically vary along a single stimulus dimension (Abraham et al., 2004; Bowman, Kording, & Gottfried, 2012; Boyle, Djordjevic, Olsson, Lundstrom, & Jones-Gotman, 2009; Kepecs, Uchida, Zariwala, & Mainen, 2008; Khan, Thattai, & Bhalla, 2008; Rinberg, Koulakov, & Gelperin, 2006; Uchida & Mainen, 2003; Wesson, Carey, Verhagen, & Wachowiak, 2008). Subjects first underwent olfactory discrimination learning where one odor mixture (CS+) was paired with either an aversive or neutral outcome, and another odor mixture (CS−) was never paired with an outcome. Subjects’ responses were then tested on a range of odor mixtures along the same perceptual dimension. We hypothesized that aversive outcomes would facilitate olfactory associative learning, and would result in a peak shift, where maximal responding would be displaced from the CS+ in the direction away from the CS−. Moreover, based on our previous findings (Li et al., 2008), we hypothesized that aversive outcomes would enhance olfactory discrimination, resulting in a steeper generalization gradient.

METHODS

Participants

Sixty-four participants provided written informed consent to take part in this behavioral experiment, which was approved by the Institutional Review Board of Northwestern University. All subjects were non-smokers under the age of 40 and reported no history of significant neurological or psychiatric conditions, significant breathing problems, or any difficulty with smell. The final sample included 30 participants. Fifteen participants (9 females; mean age ± SEM: 25.37 ± 0.88) were assigned to the aversive outcome group and received a mild electric shock as outcome, whereas another 15 participants (8 females; 25.62 ± 0.91) were assigned to the neutral outcome group, receiving an air puff as the control outcome condition. The remaining 34 participants were excluded after learning because their behavioral performance during the last conditioning block did not reach criteria.

Stimuli

Odor stimuli

Odor stimuli consisted of mixtures of two monomolecular odorants: beta-pinene (pine smell) and isoamyl acetate (banana smell). To match the two odorants at a “moderate” odor intensity level, beta-pinene was diluted 22% (0.95mL/4.25mL) in mineral oil, and isoamyl acetate was diluted 0.9% (45μL/5mL) in mineral oil. The odorants were delivered using a custom-built computer-controlled olfactometer, which permitted odor delivery without tactile or auditory confounds.

Unconditioned stimuli

The US outcomes consisted of either a mild electrical stimulation or an air puff to the left hand, serving as aversive and neutral outcomes, respectively. Electrical stimulation was administered via a single square-wave pulse of current generated by a GRASS S-48 stimulator through bipolar Ag/Au electrodes for 500 ms (Li et al., 2008). The strength of the electrical current intensity was calibrated per subject, titrated such that the stimulation was deemed ‘uncomfortable but not painful.’ This protocol was based on previously described methods from our lab (Hauner, Howard, Zelano, & Gottfried, 2013; Li et al., 2008).

Experimental procedures

We developed a novel olfactory discrimination paradigm to examine associative learning and generalization. This paradigm consisted of three different phases: an individualized odor stimulus calibration phase; a discrimination learning phase, where subjects learned to associate one odor mixture with outcome and another odor mixture with no outcome; and a generalization phase, where subjects were tested on a broad range of odor mixtures and asked to indicate if the current mixture was the one that predicted outcome, no outcome, or neither.

Odor stimulus calibration

Subjects completed a two-day experiment. On Day 1, subjects provided perceived intensity and proportion ratings for 21 binary odor mixtures ranging from 100% banana (0% pine) to 100% pine (0% banana), in 5% concentration steps. Odor intensity, as well as the perceived proportions of banana and pine smells in each mixture, were rated three times on visual analog scales (63 total trials). These perceptual ratings were used for individualized odor stimulus calibration, providing a way to accommodate variation in how subjects perceived the linear changes in the odor mixtures and how they used the rating scales. The banana:pine proportion ratings were fit to a sigmoid function to establish a final set of 11 odor mixtures, perceptually spaced in equal increments between the two pure odorants (Fig. 1B). To span the entire perceptual space, the two pure odorants were always included as anchors at the ends of the stimulus array.

Figure 1.

Experimental paradigm and odor stimulus calibration.

A. Day 1 and day 2 occurred over consecutive days. On day 1, subjects first completed baseline ratings. Next, subjects underwent an olfactory discrimination learning task, where they learned to associate one odor mixture with an outcome (either aversive or neutral) and another odor mixture with no outcome. Day 2 began with a refresh discrimination session, followed by a generalization test in extinction. Test trials assessed if subjects’ learned odor-outcome associations were generalized to odor mixtures along the same perceptual dimension.

B. Odor stimuli were individualized for each subject based on their perceived proportion ratings. The points represent 11 odor mixtures in perceptually equal increments between pure banana and pure pine (collapsed across all subjects for visualization). Error bars depict SEM.

C. Odor stimuli used during discrimination learning and generalization test phases. The CS+ and CS− mixtures were used during learning (day 1) and refresh sessions (day 2) only, while the 11 other test mixtures were used only during the generalization phase. Assignments of the CS+ and CS− stimuli to the banana-dominant and pine-dominant mixtures were counterbalanced across subjects.

After creating the 11 generalization test mixtures spaced at equal perceptual increments, two additional odor mixtures were created, serving as the CS+ and CS− (Fig. 1C). These two mixtures were positioned relatively close to each other along the perceptual gradient. The CS− and CS+ mixtures fell between mixtures 4 and 5 (35% perceived pine) and between mixtures 7 and 8 (65% perceived pine), respectively. Whether the CS+ was 65% or 35% pine was counterbalanced across subjects, controlling for the possibility that specific odor identity might drive the results.

Olfactory discrimination

Following odor stimulus calibration, subjects were assigned to one of two groups. The aversive outcome group received an electric shock as the unconditioned stimulus while the neutral outcome group received an air puff as the unconditioned stimulus. In both instances, outcomes were delivered to the left hand. In an initial phase of the experiment, subjects completed an olfactory discrimination learning task, where they learned to associate a specific odor mixture (CS+) with an outcome and another odor mixture (CS−) with no outcome (Fig. 1A). More specifically, subjects discriminated between the CS+ mixture, paired with outcome, and the CS− mixture, not paired with outcome. On each trial, subjects were cued to sniff and presented with either the CS+ or CS− odor mixture. A partial reinforcement paradigm was used, in which the CS+ mixture was repeatedly paired with outcome 75% of the time. The other 25% of CS+ trials were not paired with outcome, providing a way to analyze physiological responses to the CS+ without the confounding presence of outcome. The CS− mixture was never paired with outcome. By repeatedly pairing the CS+ odor mixture with outcome and the CS− mixture with no outcome, subjects learned to discriminate between the CS+ and CS− mixtures. Subjects had 2,500 ms from the onset of the odor mixture to respond whether the mixture was associated with outcome.

Subjects made their responses by pressing buttons corresponding to ‘yes,’ indicating the presented odor was the CS+ mixture, or ‘no,’ indicating the presented odor was the CS− mixture. Following the response time, a 1,000 ms delay occurred, after which the outcome was or was not delivered. Outcome presentation was independent of behavioral response accuracy and followed the partial reinforcement schedule described above. Behavioral responses across all phases of the study were measured in response to the odor stimuli, before the onset of the US. This allowed us to examine behavioral responses without potential confounds related to the outcome. The initial conditioning phase consisted of 80 trials, 40 of each CS. Trials were pseudorandomly delivered in 5 blocks of 16. Learning accuracy was calculated from this initial olfactory discrimination session as the sum of correctly identified CS+ trials and correctly identified CS− trials, divided by the total number of response trials for that block. Correctly identified CS+ trials were those where the CS+ mixture was presented, and subjects responded ‘yes;’ correctly identified CS− trials were those where the CS− mixture was presented, and subjects responded ‘no.’ Refresh accuracy was computed in the same way.

To ensure robust learning and performance, response accuracy criterion was set at 70% averaged across the CS+ and CS− for the last block of 16 trials. A criterion of 70% was chosen to ensure that performance would be significantly above chance (50% during discrimination learning), increasing the likelihood that learned associations would be maintained into the second day of the study. Without adequate learning of odor-outcome associations, it would not have been possible to test the extent to which outcome associations generalized to similar stimuli during the test phase. Subjects needed to achieve 70% accuracy across the last block of trials to be eligible for the remaining study phases, and those who did not reach that criterion did not participate in the remaining study sessions (n=34).

Test in extinction

On Day 2, subjects returned to complete the remaining phase of the study. The task began with a refresh discrimination session. During all refresh sessions, only the CS+ and CS− were delivered, with the same outcome contingencies as the original discrimination phase. Subjects then completed 6 test sessions, with 5 alternating refresh sessions to minimize extinction of the original odor-outcome pairing (Fig. 1A). Each test run consisted of 4 presentations of each of the 11 odor mixtures, totaling 44 trials per run. Using a wide range of odor stimuli allowed for testing the extent to which outcome associations were generalized to similar stimuli (Ghirlanda & Enquist, 2003; Hanson, 1959; Kahnt et al., 2012; Kahnt & Tobler, 2016; Spence, 1937; Thomas, 1993). The test phase took place during extinction, where outcome was never delivered, to prevent new associative learning to the test stimuli. Additionally, to further prevent new learning, the actual CS+ and CS− odor mixtures were not present during test. As in the initial discrimination session, subjects were cued to sniff and asked to indicate whether the delivered odor mixture was the CS+ mixture, associated with outcome, the CS− mixture, not associated with outcome, or a novel odor mixture, by selecting buttons associated with ‘yes’, ‘no’, or ‘don’t know.’ Subjects were instructed to select ‘don’t know’ if they believed the presented odor mixture was not previously associated with either the presence or the absence of outcome, or if they were unsure. All behavioral responses were acquired within the same 2,500 ms time window as the discrimination learning and refresh sessions. Responses where subjects identified the test stimuli as the CS+ stimulus were used as an index of olfactory generalization.

Modeling olfactory generalization gradients

Model selection and comparison with behavioral data

While the peak-shifted generalization gradients can be directly observed through subjects’ responses, the excitatory and inhibitory gradients comprising the peak-shifted gradient are only theoretical. However, using a computation modeling approach allowed us to explore these excitatory and inhibitory gradients, surrounding the CS+ and CS−, respectively (Fig. 3B). The difference between these two gradients predicts the peak-shifted responding that results in the behavioral generalization gradients (Fig. 3A).

Figure 3.

Behavioral generalization gradients.

A. Behavioral responses during the generalization test. The proportion of responses where subjects endorsed odor mixtures as the CS+ (percentage of CS+ responses) was calculated for each of the 11 odor mixtures across all test sessions. Red and blue dots denote the empirical values for each odor mixture, for shock and puff outcomes, respectively. Solid lines are the model fits to the data (as shown in B).

B. Schematic of the generalization model. Responses were modeled as reflecting the difference between an excitatory gradient (dashed line), centered on the CS+, and the inhibitory gradient (dotted line), centered on the CS−. Additionally, we modeled the minimum response value (not shown; see Methods). The difference between these excitatory and inhibitory gradients yields a peak-shifted generalization gradient (solid line)

C. Distribution of group-wise parameter differences (displayed as puff – shock). The green line represents empirical differences between the shock and air puff groups. The width of the excitatory gradient in the aversive outcome group was significantly narrower than the air puff group (permutation test, p = 0.0408). This suggests that aversive outcomes affect generalization responses by narrowing the underlying excitatory gradient.

D. Model-based excitatory and inhibitory gradients, centered at the CS+ and CS−, respectively, plotted separately for each group. The difference between the excitatory and inhibitory gradients results in the observable peak-shifted generalization gradient show in panel A. As shown in panel C, the width of the excitatory gradient was significantly narrower in the shock group, whereas the inhibitory gradient did not differ significantly between groups (p = 0.0812).

Error bars depict SEM. *p < 0.05, permutation test; n.s., non-significant.

To determine the mechanism by which outcome valence impacts generalization gradients, we modeled the excitatory and inhibitory gradients, testing two different approaches with different numbers of free parameters. First, using a same-width approach, we modeled our data by fitting two free parameters, one parameter for equal widths for excitatory and inhibitory Gaussians, and one parameter for the minimum response, corresponding to the specific mixture that subjects identified as the CS+ odor least often. Second, using a different-width approach, we modeled the data to fit three free parameters: separate widths for the excitatory and inhibitory Gaussians, centered around the CS+ and CS− values, respectively, and the minimum response parameter. Here, the model estimates the width of each Gaussian independently, allowing for flexibility in the width of generalization, where one Gaussian can be wider than the other (Fig. 3B). Next, we tested which model was a better fit to the behavioral generalization data using a leave-one-out cross-validation design. That is, the model parameters were estimated on 29 subjects and prediction performance was computed on the remaining left-out subject. Because cross-validated prediction accuracies were higher for the 3-parameter model, this model was selected for all further analyses. In addition, to test whether the fit of this model differed between groups, we computed model fits separately for each group.

Comparing generalization gradients

Next, we tested whether model parameters differed between the two groups using random permutation tests. That is, all subjects from the shock and air puff groups were pooled together and randomly assigned to two groups of 15 subjects each; the 3-parameter model was estimated in each group and the resulting parameters compared. This was repeated 5,000 times, resulting in a distribution of groupwise parameter differences that would be expected if group assignments were random. We used this distribution to determine the p-values (one-tailed) for the empirically observed parameter differences.

Respiratory and skin conductance acquisition and analysis

Physiological data were collected during all study phases to examine implicit measures of olfactory generalization. Subjects were affixed with a breathing belt around the rib cage to monitor respiration (Gottfried, Deichmann, Winston, & Dolan, 2002; Li, Luxenberg, Parrish, & Gottfried, 2006). The output of the belt was recorded using a PowerLab 8/35 data acquisition hardware and software in LabChart 7 (AD Instruments) (Li et al., 2008), and analyzed in MATLAB. Subject-specific sniff waveforms were extracted for each trial starting at sniff onset to 6 s post onset. The respiratory trace for each run was temporally smoothed using a moving window of 300 ms. Each trial-specific sniff trace was baseline corrected by subtracting the mean activity in the 500 ms window preceding sniff onset, and then sorted per odor mixture. Sniff responses were characterized by the maximum amplitude following sniff onset, to a trough following the maximum, with the starting point as the trough preceding the maximum. Any traces where no minimum preceded the maximum (e.g., a negative slope), or the maximum-minimum difference was ± 3 std away from the mean for that specific odor mixture condition, were removed and replaced by the mean of the remaining traces of the corresponding odor mixture condition. Sniff traces were then normalized within each run by subtracting the peak amplitude of all trials in the entire task run. Traces were averaged per odor mixture condition, and again defined by the maximum amplitude following sniff onset to a trough following the maximum, with the starting point as the trough preceding the maximum. For each odor mixture condition, sniff volume was calculated as the area under the curve and averaged across test runs.

Skin conductance response (SCR) was acquired from two Ag-AgCl electrodes placed on the fingers of the subject’s left hand, using PowerLab and an accompanying SCR pre-amplifier module. Analysis of SCR waveforms was conducted in MATLAB after lowpass filtering (3 Hz). Subject-specific SCR waveforms were extracted for each trial within a 14 s time window beginning at sniff onset. SCR waveforms were baseline-corrected in each trial by subtracting the mean SCR activity in the 1 s preceding sniff onset. Evoked SCR responses were sorted across each odor condition and characterized by the maximum of the SCR. In accordance with prior methods (Hauner et al., 2013; Li et al., 2008), only trials with an evoked deflection were included in the SCR analysis. All SCR trials with a smaller deflection than 0.01 μS, or those with a peak-trough difference ± 4 std away from the mean for that specific odor condition, were removed and replaced by the mean of the remaining waveforms of that odor condition, similar to the respiration analysis above. SCR traces were then normalized within each run by subtracting the peak SCR amplitude for all trials in the entire task run. All traces were averaged per odor mixture condition. Odor evoked SCR traces were defined by the maximum amplitude following sniff onset to a trough following the maximum, with the starting point as the trough preceding the maximum. For each odor mixture condition, sniff volume was calculated as the area under the curve and averaged across test runs.

Statistics

Statistical tests for behavioral and physiological measures were established using two-way mixed ANOVAs, one-sample, and two-sample t-tests, as appropriate. All reported p-values are two-tailed, unless otherwise noted. Error bars were adjusted for between-subjects differences by subtracting the mean across all conditions for each subject (Antony, Gobel, O’Hare, Reber, & Paller, 2012; Cousineau, 2005). Statistical analyses were conducted using MATLAB (MathWorks).

RESULTS

Odor-outcome associations during discrimination training

Outcome valence impacts associative learning

Subjects underwent olfactory discrimination training to learn specific odor-outcome associations (Fig. 1A). Conditioned stimuli included mixtures of banana and pine odors, with the CS+ and CS− stimuli counterbalanced across subjects. Two groups of participants completed this paradigm. One group received a CS+ paired with mild electric shock, as the aversive outcome, while the other group received a CS+ paired with an air puff, as the neutral outcome. In both groups, the CS− odor was never paired with an outcome. During the initial training session, subjects had to discriminate between the CS+ and CS− odor mixtures. Both groups showed an increase in learning performance from the first block of trials to the last, with the shock group increasing from 68.87% to 78.33%, and the air puff group increasing from 57.11% to 66.04% (Fig. 2A). Across 64 subjects, a two-way ANOVA on discrimination accuracy, with factors of outcome group (shock or air puff) and learning block, showed a robust increase in discrimination performance over time (main effect of time: F_(4,248) = 8.0904, p = 3.8019 × 10⁻⁶), and established that subjects learned the cue-outcome associations. Importantly, aversive outcomes facilitated associative learning over and above the neutral outcomes (main effect of group: F_(1,62) = 14.492, p = 0.0003; group × time interaction: F_(4,248) = 0.3111, p = 0.8703). Likewise, the aversive outcome resulted in significantly higher accuracy for the last block of training when compared to the neutral outcome (t₆₂ = 3.0192, p = 0.0037, two-sample t-test), highlighting valence-dependent learning differences between groups. In examining the data at the single-subject level, it was clear that many subjects failed to meet learning criterion of at least 70% for the last block of 16 trials. In the shock group, only 1 subject (out of 16 subjects) failed to meet criteria, whereas in the air puff group, 33 of 48 subjects were unable to meet these same criteria (chi-square = 18.8235, p = 1.4339 × 10⁻⁵). Subjects who did not demonstrate learning to criteria were excluded after discrimination training.

Figure 2.

Behavioral performance during discrimination learning.

A. Learning was measured as the percentage of correctly identified CS+ and CS− trials. Across all subjects (shock n = 16; air puff n = 48), the aversive outcome resulted in significantly higher learning accuracy for the last block of training when compared to the neutral outcome (t₆₂ = 3.0192, p = 0.0037, two-sample t-test). Accuracy criterion was set at 70% for the last block (dashed line).

B. Subjects in both groups (shock n = 15; air puff n = 15) learned the odor-outcome associations to criterion by the last block of trials. Importantly, there were no differences between groups in discrimination accuracy during the last block of learning (t₂₈ = −0.5971; p = 0.5552; two-sample t-test).

Error bars depict SEM. *p < 0.005, two-tailed t-test; n.s., non-significant.

No differences in learning between groups matched for performance

Among the remaining 30 subjects (15 in the shock group, 15 in the air puff group), learned associations between odor mixtures and outcomes of different valences at comparable levels during discrimination training (Fig. 2B). Discrimination accuracy increased for both groups across time (two-way mixed ANOVA; main effect of time: F_(4,112) = 12.526, p = 1.8628 × 10⁻⁸; shock group increased from 70.38% in the first block of trials to 79.80% in the last block; air puff group increased from 59.38% in first block of trials to 81.81% in the last block), but there were no differences between groups (main effect of group: F_(1,28) = 2.277 p = 0.1425; group × time interaction: F_(4,112) = 2.050, p = 0.092). Most importantly, for the last block of learning, we found no difference between groups (t₂₈ = −0.5971; p = 0.5552; two-sample t-test). Confirming that the two groups learned odor-outcome associations equally well ensures that any observed group differences in behavioral or physiological measures resulted from the impact of outcome valence on generalization at test, as opposed to differences in task difficulty per se.

Olfactory generalization

Behavioral generalization

On the second day of the study, subjects took part in the generalization test (Fig. 1A). This phase of the experiment was done in extinction, that is, without presentation of the outcomes, to prevent the formation of new associations between the larger set of odor mixtures and the outcomes, which might complicate data interpretation. An array of 11 odor mixtures, perceptually spaced in equal increments between pure banana and pure pine, comprised the stimulus set (Fig. 1C), and on each trial, one of the 11 mixtures was delivered to the subject. During the task, subjects were asked to respond whether the presented mixture was either the CS+ odor or the CS− odor, or a novel mixture not previously encountered in the day 1 learning session. To examine generalization behavior, we calculated the proportion of responses where subjects identified an odor mixture as the CS+ (percentage of CS+ responses) for each of the 11 odor mixtures across all test sessions.

Behavioral responses to test stimuli in both groups resulted in peak-shifted generalization gradients, with increased responding to stimuli displaced away from the CS+, in the direction away from the CS− (Fig. 3A). To quantify this, we tested for asymmetry surrounding the CS+ mixture by comparing responses to the two mixtures directly right of the CS+ mixture to the two mixtures directly left of the CS+ mixture. In both groups, responses were significantly greater to odors right of the CS+ (comparison of right vs. left of the CS+, shock group: t₁₄ = 3.83, p = 0.0018; air puff group: t₁₄ = 3.87, p = 0.0017; paired t-test). This demonstrates that the generalization gradient was asymmetric around the CS+, hence revealing a peak shift. If the peak response had actually been centered at the CS+ mixture, there would have been no difference between mixtures on the right and left of the CS+.

Importantly, behavioral responses to test stimuli differed between groups. A two-way mixed ANOVA on the percentage of CS+ responses during test revealed a main effect of odor mixture on subjects’ generalization responses (F_(10,280) = 73.384, p = 2.78 × 10⁻⁷²), no effect of group (F_(1,28) = 0.6503, p = 0.4268), but a significant interaction between group and odor mixture (F_(10,280) = 1.9161, p = 0.0429). Relative to neutral outcomes, aversive outcomes resulted in a steeper response gradient. Post-hoc comparisons between the slopes of the gradients revealed a significantly steeper slope in the shock group compared to the air puff group (shock mean: 8.49 ± 0.90 SEM; air puff mean: 6.14 ± 0.79; t₂₈ = 1.95; p = 0.030, one-tailed). The intercept was not significantly different between groups (t₂₈ = −1.64; p = 0.112). These findings demonstrate that outcome valence altered generalization behavior.

To determine whether discrimination refresh sessions served as potential reminders of the learned odor-outcome associations, secondarily impacting generalization behavior, we compared responses across the six generalization runs. Specifically, we estimated the linear change of responding across generalization test runs for each odor mixture and subject. A two-way (group by mixture) ANOVA on these slopes revealed no main effect of odor mixture on subjects’ generalization responses (F_(10,280) = 0.676, p = 0.746), no main effect of group (F_(1,28) = 0.268, p = 0.609), and no significant group by odor mixture interaction (F_(10,280) = 1.085, p = 0.374). This suggests that discrimination refresh sessions did not influence generalization, although as “null results,” it is possible that with greater numbers of subjects, the refresh reminders might have had some influence.

Aversive outcomes narrow excitatory gradients

The behavioral responses reported above suggest that outcome valence alters generalization behavior. Assuming a model in which generalization is driven by a difference between excitatory and inhibitory gradients around the CS+ and CS−, respectively (Ghirlanda & Enquist, 2003; Pearce et al., 2008; Spence, 1937), these group differences could be driven by either or both of these gradients. Of note, the excitatory and inhibitory gradients cannot be directly observed from behavior, but they can be estimated using a computational model. To this end, we modeled behavioral responses during test trials using a combination of excitatory and inhibitory Gaussian gradients around the CS+ and CS−, respectively (Fig. 3B).

We first compared two different models with different numbers of free parameters. Specifically, we compared the performance of a model with three free parameters (excitatory gradient, inhibitory gradient, and minimum response), and a simpler model with two free parameters (equal widths for excitatory and inhibitory gradients, minimum response; see Methods). A leave-one-subject-out cross-validation analysis (see Methods) revealed that the 3-parameter model predicted behavior better than the 2-parameter model.

Next, we computed how well the 3-parameter model was able to predict the generalization responses for each subject. Across odor mixtures and subjects, we found a significant correlation between the responses of the model and subjects’ responses in both groups (shock group r = 0.8033, p < 0.001; air puff group r = 0.7814, p < 0.001). Importantly, these correlations were not significantly different between groups (Z = 0.53, p = 0.298), indicating that the model fit did not differ between groups. Figure 3A shows the behavioral responses to each odor mixture, along with the model fits for each group, illustrating that the model reproduced behavioral responses in each group accurately, including the overall width of the generalization gradients and the height of the peaks.

In a final step, we tested whether model parameters differed between the two groups. This analysis aimed to uncover whether differences in generalization responses were driven by differences in the inhibitory gradient around the CS−, the excitatory gradient around the CS+, or both. The width of the excitatory gradient in the shock group was significantly narrower than in the air puff group (permutation test, p = 0.0408), suggesting that aversive outcomes specifically narrowed excitatory generalization gradients. The width of the inhibitory gradient was also nominally narrower in the shock group, but this difference was not significant (permutation test, p = 0.0812). There was no significant difference between groups for the minimum response parameter (permutation test, p = 0.6508). The two model-based gradients are plotted separately for each group in Figure 3D. These findings provide evidence that outcome valence affects generalization behavior by narrowing the width of the (hidden) excitatory generalization gradient around the CS+.

Physiological measures of generalization

Respiratory and skin conductance responses were also monitored throughout the experiment as indirect measures of olfactory generalization. Sniff-related measures and odor-evoked skin conductance response (SCR) waveforms were extracted for each odor mixture from all test runs. Here, the prediction was that physiological measures may exhibit generalization gradients similar to the behavioral gradient. A two-way mixed ANOVA on sniff volume with factors of group and mixture revealed significant differences across odor mixtures (main effect of odor mixture: F_(10,280) = 6.302, p = 1.010 × 10⁻⁸), differences between the shock and air puff groups (main effect of group: F_(1,28) = 4.29, p = 0.0476), and a significant group × odor interaction (F_(10,280) = 2.132, p = 0.0222; Fig. 4B). A two-way mixed ANOVA on SCR volume revealed differences across odor mixtures (main effect of odor mixture: F_(10,280) = 1.9473, p = 0.0391), but no group differences (main effect of group: F_(1,28) = 1.2824, p = 0.2824) and no significant group × mixture interaction (F_(10,280) = 0.4302, p = 0.9312; Fig. 4D). These physiological generalization gradients indicate that aversive outcomes resulted in increased sniff volume, specifically to odor mixtures that are dissimilar to the CS+ in the direction away from the CS−, suggesting that respiratory generalization coincides with behavioral generalization. It is worth noting that while unpleasant or aversive odors per se often induce smaller and shorter sniffs (Arzi et al., 2012; Bensafi et al., 2014; Mainland & Sobel, 2006), in the setting of our study, where task demands required subjects to discriminate between similar odors, the observation of larger sniff volumes would be consistent with the strategy of making larger sniffs to gain as much information as possible before making a decision.

Figure 4.

Physiological generalization.

A. The averaged respiratory trace across all test sessions, time-locked to sniff cue onset, for the aversive outcome group (red) and neutral outcome group (blue).

B. Average sniff volume as a function of test odor mixture showed a main effect of odor mixture (F_(10,280) = 6.302, p = 1.010 × 10⁻⁸), a main effect of group (F_(1,28) = 4.29, p = 0.0476), and a group × odor interaction (F_(10,280) = 2.132, p = 0.0222).

C. The averaged skin conductance response (SCR) across all test sessions, time-locked to sniff cue onset, for the aversive outcome group (red) and neutral outcome group (blue).

D. Average SCR volume generalization across odor mixtures showed a main effect of odor mixture (F_(10,280) = 1.9473, p = 0.0391), but no effect of group (F_(1,28) = 1.2824, p = 0.2824) and no interaction (F_(10,280) = 0.4302, p = 0.9312).

Traces are shown as mean (line) and SEM (shaded areas). Error bars depict SEM.

Behavioral differences are not driven by discrimination performance

To ensure participants maintained the odor-outcome associations during the entirety of the task, did not habituate to the odors, and were continuously able to discriminate between the CS+ and CS− mixtures, performance accuracy was measured across refresh runs. A two-way mixed ANOVA on discrimination performance in refresh runs showed no significant differences between groups (F_(1,28) = 3.328, p = 0.078), no main effect of time on discrimination accuracy (F_(1,28) = 0.484, p = 0.787), and no significant group × time interaction (F_(5,140) = 1.8947, p = 0.098; between-subject factor of group and within-subject factor of time). It is therefore unlikely that group discrimination performance on the second day can account for behavioral generalization differences.

Valence of conditioned stimulus did not influence outcome valence

Given the affective nature of odors, it is possible that the valence of the CS mixture itself influenced the perceived valence of outcome. To test for such effects, we split our groups not by outcome valence, but by whether the CS+ was more pine (65% perceived pine) or more banana (35% perceived pine). The prediction here was that if CS+ valence affected responding, we would see a significant effect of group or group by mixture interaction. A two-way mixed effects ANOVA on percentage of CS+ responses revealed a main effect of odor mixture on generalization responses (F_(10,280) = 74.265, p = 8.69 × 10⁻⁷³), but critically no main effect of group (F_(1,28) = 0.063, p = 0.803), and no significant interaction between group and odor mixture (F_(10,280) = 1.847, p = 0.053).

DISCUSSION

The ability to learn associations between sensory stimuli and outcomes, and to generalize these associations to novel stimuli, is a fundamental component of adaptive behavior (Bouton, 2006; Guttman & Kalish, 1956; McLaren & Mackintosh, 2002; Pearce, 1987; Shepard, 1987; Spence, 1937). Here, using a novel olfactory discrimination learning paradigm, we tested how outcome valence impacts associative learning and generalization in humans. We found that discrimination learning was enhanced for odors associated with aversive compared to neutral outcomes. Moreover, we found that generalization gradients were affected by outcome valence, such that excitatory gradients were narrower for aversive outcomes. These results demonstrate that aversive outcomes enhance learning and discriminatory responses during generalization.

Our results are in line with previous studies showing that aversive outcomes enhance learning compared to neutral outcomes. Previous work has shown that outcome valence plays an important role in associative learning and generalization (Dunsmoor et al., 2014; Dunsmoor et al., 2017; Dunsmoor & Murphy, 2014; Schechtman et al., 2010; Staddon, 1983). The impact of physical intensity (strength) of the outcome has been established in associative learning models (Pearce & Hall, 1980; Rescorla, 1972; Wagner, 1981), where more intense outcomes lead to stronger conditioning (Gottfried & Dolan, 2004). Rodent fear conditioning studies have validated the role of outcome by varying foot-shock intensity to show that stronger foot-shocks give rise to stronger fear learning (Ader, 1972; Annau & Kamin, 1961; Baldi et al., 2004; Cordero, Merino, & Sandi, 1998; Davis & Astrachan, 1978; Phillips & LeDoux, 1992).

In addition to impacting learning, outcomes of different valences also alter patterns of generalization, even when matched for learning performance. More specifically, studies show that aversive reinforcers induce wider generalization gradients than positive reinforcers, such that more stimuli evoke behavioral responses similar to the CS+ (Dunsmoor et al., 2014; Dunsmoor & Murphy, 2014; Resnik et al., 2011; Schechtman et al., 2010). It has been suggested that aversive outcomes modulate perceptual discrimination thresholds, thereby impairing discrimination and widening generalization (Dunsmoor & LaBar, 2013; Laufer & Paz, 2012; Resnik et al., 2011; Shalev et al., 2018). However, in contrast to these findings, our results suggest aversive outcomes result in narrower, rather than wider, generalization gradients. This may reflect difference in modalities, and is in line with previous work showing that aversive outcomes enhance olfactory discrimination, such that initially indistinguishable odors become discriminable after aversive conditioning (Ahs, Miller, Gordon, & Lundstrom, 2013; Li et al., 2008). Thus, our results may highlight differences between different sensory modalities.

Our findings could be seen as being counter-intuitive, as it might seem practical to generalize across similar stimuli that predict aversive outcomes. However, improved discrimination could allow for more precise identification of predictors of aversive outcomes. This is particularly relevant for the olfactory system, given the extensive range of perceptually similar odors that may predict vastly different outcomes. It has been shown that the perceptual features of odors from a single source may vary extensively over time (Cain, 1992; Polak, 1973), suggesting that olfactory stimuli are inherently less stable than their auditory or visual counterparts, making them more amenable to associative plasticity. In this context, it reasonably follows that stimuli paired with aversive outcomes would be better discriminated than those paired with neutral outcomes. As discussed above, it has been previously shown that associative learning with aversive outcomes improves rather than impairs olfactory discrimination (Ahs et al., 2013; Li et al., 2008). Generalization may be an active process where learned associations are transferred to similar stimuli despite the ability to detect perceptual differences between stimuli (Dunsmoor & Paz, 2015; Guttman & Kalish, 1956; Kahnt & Tobler, 2016; Shepard, 1987). Our present results suggest that olfactory generalization is not passively driven by perceptual features of odors, but an active inference process centered around outcome valence.

Our two-day study design was closely modeled to the design of Kahnt and Tobler, 2016, where the generalization test phase was conducted on a separate day. By separating the initial odor calibration and discrimination sessions from the generalization and refresh sessions, we hoped to limit task fatigue and olfactory habituation during the three-hour paradigm. Based on previous studies that have combined learning and generalization sessions in a single day (Kahnt et al., 2012), and those showing peak shift effects after a 24-hour delay between training and test (Wisniewski et al., 2009), we do not believe our findings would significantly differ if all phases of the study were conducted in a single day. That said, it is worth considering that the narrowing of generalization gradients in the aversive outcome group may have been influenced by consolidation mechanisms, whereby a night of sleep may have resulted in consolidation effects that could help account for our counterintuitive findings.

Another key feature of our experimental design is that the conditioned stimuli were not presented during the probe test. This experiment was modeled after previous generalization experiments (Kahnt et al., 2012; Kahnt & Tobler, 2016), which also did not include the CS+ and CS− stimuli during the generalization test. This design allows us to compare responses to a range of test stimuli that have not been paired with an outcome. Therefore, and critically, differences in responding to the test stimuli cannot be explained by a differential history of experience with stimulus-outcome associations per se, and rather can only be explained by differences in the similarity of test stimuli to the original CS. Probe tests in generalization tasks are routinely performed under extinction conditions (i.e., in the absence of any outcome) (Ghirlanda & Enquist, 2007; Kahnt et al., 2012; Kahnt & Tobler, 2016; Purtle, 1973), because including the original CSs would alter their original associative strengths. Thus, by excluding the CS+ and CS− from the test phase, we were able to examine how learned associations generalize to similar test stimuli, without the potential confounds of altering the originally learned stimulus-outcome associations. Additionally, because the CSs were very close to the two adjacent test stimuli, it is unlikely that subjects would have been able to reliably discriminate between the CS+ (or the CS−) and the two test stimuli closest (‘left’ and ‘right’) to it. As such, it is reasonable to infer that subjects perceived the CS+ (and the CS−) effectively the same as the two adjacent test stimuli.

Just as generalization involving appetitive and aversive outcomes may involve different mechanisms (Schechtman et al., 2010), it is worth considering that different sensory modalities also engage different neural mechanisms. Previous studies have shown that an active decision-making process exists for finer discrimination and is driven by cortical networks (Li et al., 2008; Weinberger, 2007). Insofar as our findings suggest that olfactory generalization is an active process, a key component of this process may involve orbitofrontal cortex (OFC), which is thought to play a key role in model-based inference (J. L. Jones et al., 2012; Wang, Schoenbaum, & Kahnt, 2020). Another possible brain area involved may be the hippocampus, as studies in the visual system have found that hippocampal-striatal (Kahnt et al., 2012) and hippocampal-midbrain (Kahnt & Tobler, 2016) connectivity flexibly modulates the extent of generalization. Aversive discrimination learning and generalization are also influenced by amygdala networks (Armony, Servan-Schreiber, Romanski, Cohen, & LeDoux, 1997; Chavez, McGaugh, & Weinberger, 2009; Dunsmoor, Prince, Murty, Kragel, & LaBar, 2011; Laxmi et al., 2003). Lastly, the piriform cortex may also be involved, as previous findings have shown that aversive learning induces olfactory cortical plasticity that corresponds with discrimination performance (S. V. Jones, Stanek-Rattiner, Davis, & Ressler, 2007; Li et al., 2008; Weinberger, 2004). While speculative, these considerations may help guide future hypotheses for elucidating the neural processes underlying olfactory generalization and discrimination, two important sides of the same olfactory coin, but with very divergent consequences.