Task reward structure shapes rapid receptive field plasticity in auditory cortex

Abstract

As sensory stimuli and behavioral demands change, the attentive brain quickly identifies task-relevant stimuli and associates them with appropriate motor responses. The effects of attention on sensory processing vary across task paradigms, suggesting that the brain may use multiple strategies and mechanisms to highlight attended stimuli and link them to motor action. To better understand factors that contribute to these variable effects, we studied sensory representations in primary auditory cortex (A1) during two instrumental tasks that shared the same auditory discrimination but required different behavioral responses, either approach or avoidance. In the approach task, ferrets were rewarded for licking a spout when they heard a target tone amid a sequence of reference noise sounds. In the avoidance task, they were punished unless they inhibited licking to the target. To explore how these changes in task reward structure influenced attention-driven rapid plasticity in A1, we measured changes in sensory neural responses during behavior. Responses to the target changed selectively during both tasks but did so with opposite sign. Despite the differences in sign, both effects were consistent with a general neural coding strategy that maximizes discriminability between sound classes. The dependence of the direction of plasticity on task suggests that representations in A1 change not only to sharpen representations of task-relevant stimuli but also to amplify responses to stimuli that signal aversive outcomes and lead to behavioral inhibition. Thus, top-down control of sensory processing can be shaped by task reward structure in addition to the required sensory discrimination.

As we interact with a world in flux, our brains adjust their responses to sensory stimuli, allowing us to meet changing behavioral demands (1, 2). Numerous studies have shown that attention contributes to this process by selectively modulating neural activity in brain areas that process sensory information, improving stimulus discriminability for grouping into task-relevant categories (3–13). However, the effects observed across behavioral paradigms are diverse, including changes in gain (4, 5, 7, 14), selectivity (8, 11, 15, 16), and functional connectivity (12), suggesting that the brain may use many possible strategies and mechanisms to highlight relevant stimuli and produce appropriate sensorimotor transformations. The specific effects of attention and other learned behaviors on representations, therefore, may depend not only on the required sensory grouping but also on a host of control signals reflecting task structure (17), motor responses (18, 19), associated reward (20), difficulty (21, 22), and timing of decisions (23).

To directly explore the influence of task structure on sensory representations, we recorded the activity of single neurons in primary auditory cortex (A1) during two different instrumental behaviors that required discrimination between the same two acoustic categories, pure tones and broadband noise. Both tasks had a “go/no-go” structure. The first task required a target approach behavior: subjects received a positive reward (water) for withholding a licking response during a sequence of reference noise sounds and responding by licking when a target tone occurred (24). The second task required target avoidance: subjects could freely lick water from a spout during reference sounds but received a negative reward (a mild shock) for failing to inhibit licking immediately after the occurrence of the target tone (8, 25). Thus, for identical reference and target sounds, both the learned go vs. no-go association and the reward valence associated with stimuli were reversed between the two behavior paradigms. Our use of rippled noise as reference sounds enabled us to characterize neuronal receptive field selectivity during task performance (26). Changes in selectivity between behaving and passive nonbehaving conditions reflected task-dependent changes in auditory representation (8, 22, 27).

One might predict two different outcomes during the two behaviors. If behavioral control of A1 reflects contrast between reference and target sounds, independent of task structure, one would expect to find enhanced responses to targets for both approach and avoidance tasks (8). If, however, top-down control signals reflect stimulus valence or the associated behavior (28, 29), one would expect effects during the approach task to have an opposite sign to those observed during the avoidance task. The results reported here strongly support the latter prediction, as we found that the sign of rapid behaviorally driven changes in A1 neurons depended qualitatively on the task.

Results

To compare the effects of approach and avoidance behavior on representations in auditory cortex, ferrets were trained to perform two different tasks, each of which required discrimination between the same reference and target sound classes (Fig. 1A). References were drawn from a set of 30 broadband rippled noise samples (26), and the target was a pure tone with frequency fixed during each behavioral session but varied over 5 octaves between sessions. During the approach task, animals signaled target recognition by licking a water spout. Early responses counted as false alarms and resulted in a timeout. Thus, animals withheld licking until target onset, after which they increased their lick rate to indicate recognition of the reward signal (Fig. 1B). Water was delivered from the spout only after a correct response. During the avoidance task, animals could drink from the water spout during references, and signaled target recognition by ceasing to lick. Failure to stop licking after target offset resulted in a mild tail shock. Animals performing the avoidance task maintained a high lick rate during reference sounds (Fig. 1C). They were required to stop licking only after target offset, but animals naturally adopted a strategy in which they decreased licking after target onset. Hence the onset of target tones triggered behavioral responses in both tasks, but the behavior differed. During the approach task, the target sound indicated that a response (licking) would be rewarded (the “go” stimulus), whereas during avoidance, the same target sound indicated that ongoing behavior (licking) should be inhibited (the “no-go” stimulus).

Fig. 1.

Approach vs. avoidance behaviors. (A) In both tasks, subjects were required to detect a pure tone target (red) after a random number of reference noise sounds (blue). During the approach behavior (timeline, Lower), subjects were positively rewarded with water for licking a water spout 0.1–1.0 s after target onset (green bar) and punished with a timeout for licking earlier (red bar). During avoidance, subjects were rewarded by licking a continuously flowing stream of water during the references and punished with a mild tail shock if they failed to cease licking 0.4 s after target offset. (B) Average behavior during approach experiments, plotted as a function of time after reference (blue) or target onset (red). Shading indicates one SE across sessions. Dashed lines indicate stimulus onset and offset. Licking was minimal during references, because these trials were punished as false alarms; it substantially increased after target onset, and these trials were rewarded as hits. A stereotyped lick rate (five licks per second) is reflected by multiple peaks in the target curve. (C) Average behavior during avoidance experiments, plotted as in B. During references, animals maintained an elevated lick rate to retrieve reward. Licking was attenuated 0.2–0.4 s after target onset until after offset, when licking resulted in punishment for a miss.

Opposite Tuning Changes During Approach and Avoidance.

We recorded activity from single A1 neurons during either approach (n = 270, six animals) or avoidance behavior (n = 247, eight animals). All analyzed units were recorded during experiments that met behavioral performance criteria (Fig. S1) and produced responses that met auditory signal-to-noise criteria. Task-dependent changes in neural response properties were measured by comparing sound features that evoked or suppressed neural firing during behavior vs. passive presentation of task stimuli. To characterize neuronal tuning, we performed reverse correlation between the spectrogram of the reference sounds and neural firing rate over time (26). The resulting spectrotemporal receptive field (STRF; Fig. 2) represented neural selectivity as a function of frequency and time after stimulus. Differences between STRFs estimated during behavior and passive stimulus presentation reflected task-dependent changes in spectrotemporal tuning (8, 22, 27).

Fig. 2.

Examples of spectrotemporal tuning changes during target approach and avoidance behaviors. (A) Data from one neuron during the approach task. The STRF estimated from reference sounds during passive listening (Left) indicates stimulus frequencies and time lags correlated with excited (red) or suppressed (blue) neural spiking. This neuron was excited broadly by 8,000- to 16,000-Hz stimuli, which overlapped the target tone (11,046 Hz). During behavior (Center), a notch appeared at the target frequency in the excitatory region, and the difference between the active and passive STRFs (Right) shows a 23% decrease in gain at the target frequency (“X”). (B) STRF change for a second neuron during approach, plotted as in A. Here the target frequency (11,300 Hz) overlapped an inhibitory subregion of the passive STRF. During behavior, the inhibition grew stronger, producing a net 7% decrease at the target frequency. (C) STRF change for a neuron during avoidance behavior, with the target positioned on the shoulder of an excitatory region of the passive STRF (1,250 Hz). The STRF showed a selective 10% increase at the target frequency during behavior. (D) Data from a second neuron recorded during avoidance, with the target positioned over an inhibitory region (1,350 Hz). The inhibition was mostly abolished during behavior, a 30% increase at the target frequency.

During the approach task, responses to the target frequency tended to decrease. In one neuron, the target tone—the go stimulus for this behavior—had a frequency of 11,046 Hz, which lay within the excitatory region of the STRF measured during passive listening (Fig. 2A, Left, arrow). While the animal performed the task, the excitatory region of the STRF split into two separate lobes (Fig. 2A, Center), reflecting a decrease in responsiveness to stimuli at the target frequency. Individual neurons sometimes showed STRF changes at other frequencies, but the consistent effect was a decrease at the target frequency. In a second neuron (Fig. 2B), the target frequency was 11,300 Hz, positioned within a weakly inhibitory region of the STRF. During approach behavior, this neuron showed greater excitation at frequencies away from the target. At the target frequency, however, the STRF showed increased inhibition, and thus a net decrease in its response.

The effects of avoidance behavior, for which the target was now the no-go stimulus, followed an opposite pattern (confirming earlier results) (8). In an example neuron from this data set, the STRF showed weak excitation at the target frequency during passive stimulation (1,250 Hz; Fig. 2C). During behavior, the STRF increased its responsiveness at the target frequency. In a second example (Fig. 2D), the target was positioned over an inhibitory region of the passively measured STRF. For this neuron, the inhibitory region disappeared during behavior, again a net increase at the target frequency.

We measured the average effect of approach behavior in A1 by aligning the STRF difference (behavior minus passive listening) for each neuron at its target frequency (Fig. 3A). The average STRF difference revealed a consistent decrease in responsiveness at the target frequency. Approximately one-third (86/270) of neurons underwent significant STRF changes at the target frequency during behavior (P < 0.05). The majority of these neurons (69/86) decreased their response, and the mean change across significant neurons was −24% of peak STRF amplitude (P < 0.0001). In contrast, the average STRF difference during avoidance behavior showed an opposite, selective enhancement at the target frequency (Fig. 3B). In this case, 35% (87/247) of neurons underwent significant changes at the target frequency (P < 0.05). The majority (62/87) increased their response, changing by a mean of +20% (P < 0.0001). Thus, changing the required behavioral target response (approach vs. avoidance) had a direct and opposite effect on the average sign of changes in neural responses during discrimination between the same reference and target sounds. In both cases, A1 neurons showed increased responses to no-go stimuli relative to go stimuli.

Fig. 3.

Opposite changes in spectrotemporal tuning during approach and avoidance. (A) Average STRF difference between approach behavior and passive listening, aligned at the target frequency and averaged over neurons (Left; n = 270). Red regions indicate frequencies and time lags with increased responsiveness, and blue regions indicate a decrease. The average difference shows a selective decrease in amplitude within 0.25 octave of the target frequency. The histogram (Right) plots the fraction difference at target frequency between each behaving and passive STRF. Filled bars indicate the 86 units showing significant changes (P < 0.05). The mean change across significantly modulated neurons was −24% of peak passive STRF amplitude (median −30%, both significant, P < 0.0001). (B) Average STRF difference for avoidance behavior, plotted as in A (n = 247). In this case, average amplitude increased near the target frequency. The mean change across the 87 significantly modulated neurons was a 20% increase (median 23%, both significant, P < 0.0001).

Because the changes in STRFs appeared rapidly after the onset of behavior, we asked whether they were reduced after behavior ceased. We compared STRFs measured during passive listening before and after behavior and found that, on average, the changes at target frequency diminished and STRFs returned to their prebehavioral state (Fig. 4A), although receptive field changes did persist for some neurons following the avoidance task (8, 30). Thus, in both task conditions STRF changes were largely reversed after behavior was complete.

Fig. 4.

(A) Average STRF change at target frequency during behavior (black bars) and postbehavior (white bars), relative to prebehavior baseline. The opposite sign changes for approach and avoidance largely reversed after behavior was complete. Data are shown for significantly modulated neurons with passive data recorded both before and after behavior (n = 86 approach, 59 avoidance). The gray bar plots the average STRF change during avoidance, after excluding data 150 ms before and after lick events. Error bars indicate 1 SE across neurons. (B) Average STRF change at target, grouped by the difference between neural BF and target frequency. During approach (blue, n = 86), STRF changes tended to be negative for all neurons, but the greatest decrease occurred when BF was within 0.1 octave of the target frequency. During avoidance (red, n = 87), STRF changes were positive for all BF-target frequency distances, and the magnitude was also greatest within 0.1 octave of target.

One difference between tasks was that in the avoidance behavior, animals licked a water spout during the reference sounds used for STRF estimation. Licking can produce sound and may attenuate auditory responses through the middle ear reflex. Because target frequency and neural frequency selectivity varied between experiments, it is unlikely that licking could produce the observed frequency-specific effects. However, to control for possible contributions of licking to STRF changes, we recomputed STRFs for the avoidance task after excluding data 150 ms before and after lick detection (average lick duration was 200 ms). This exclusion did not significantly affect the mean change at target frequency (Fig. 4A). We also tested for a direct correlation between licking and spiking (31) and found marginally significant motor activity in a few neurons (approach 4/86, avoidance 6/87, P < 0.05). Thus, the STRF changes during behavior cannot be attributed to peripheral motor signals.

The effects were consistent across animals (Fig. S2) and were larger during blocks with better performance (Fig. S3), indicating that the sign of receptive field changes was not random but was rather a consequence of task-imposed constraints. It is possible that long-term plasticity, reflecting differences in training (32, 33), could influence rapid task-dependent effects, but we did not observe a difference in the plasticity for animals trained on one or two tasks.

Previous studies of rapid (22) and long-term plasticity (34) have reported a relationship between baseline tuning and task effects in A1. To test for this relationship, we grouped neurons by the distance between their best frequency (BF; peak frequency of the passive STRF) and the target frequency used in the experiment (Fig. 4B). During approach, the largest STRF decreases were observed for neurons with BF within 0.1 octave of the target. During avoidance, the increase at target frequency was also most prominent for neurons with BF near the target, although the trend persisted up to 1 octave away. Thus, in both paradigms behavioral effects were most pronounced in neurons with BF near the target frequency (22, 34).

Enhanced Response to No-Go Stimuli During both Behaviors.

The preceding analysis of STRF changes focused on activity during the presentation of reference sounds. However, these results predict changes in target responses, namely that they should decrease during the approach task and increase during the avoidance task. We verified this prediction by measuring the change in the raw spike rate response to reference and target sounds between passive listening and behavior (14). We again grouped neurons according to the distance between BF and target frequency. During approach (Fig. 5A), neurons with BF within 0.1 octave of the target showed an average decrease of 14% in their target response (P < 0.01). Neurons with BF further from the target frequency showed little or no average change. In contrast, reference responses tended to increase, regardless of BF. Thus, as suggested by the average STRF change (Fig. 3A), the net effect of the approach behavior was to increase the relative response to reference, no-go sounds over target, go sounds.

Fig. 5.

Emergent representation of sound class during behavior. (A) Average fraction change in raw neural response to target (black) and reference (gray) during approach behavior, grouped by BF-target frequency distance (n = 270 neurons with both reference and target data collected during passive listening and behavior). Neurons with BF within 0.1 octave of target frequency showed decreased target vs. reference responses (*P < 0.01), whereas other neurons showed no consistent target response change. Reference responses tended to increase, regardless of BF. (B) Change in raw response during avoidance, plotted as in A (n = 174). Average target responses increased for BF within 0.1 octave of target, whereas reference responses tended to decrease for all neurons. Changes in both A and B are consistent with enhanced reference-target discriminability. (C) Performance of a linear decoder trained to discriminate reference and target sounds from neural responses during approach behavior (red) and passive listening (green, n = 270). Crosses indicate average fraction of correct classifications as a function of the number of neurons in the decoder, fit by a decaying exponential (dashed lines). On average, 11.2 neurons were required to achieve 90% accuracy during behavior, and 16.0 were required during passive listening (bars, Lower, P < 0.001). (D) Performance of a linear decoder trained on avoidance data, plotted as in C (n = 151). During behavior, an average of 13.6 neurons was required to achieve 90% accuracy, and an average of 19.2 was required during passive listening (P < 0.001).

When we compared changes in raw neural responses during avoidance, we observed a different effect (Fig. 5B). For neurons with BF near the target frequency, average target response increased (17%, P < 0.01). However, reference responses were slightly reduced. Across the entire A1 population, these changes would produce the increase at target frequency in the STRF (Fig. 3B) and, as in the case of the approach behavior, an overall increase in response to no-go sounds over go sounds.

Emergent Grouping of Reference and Target Classes During both Behaviors.

The changes in STRFs during the two behaviors suggest qualitative differences in top-down feedback to A1, but a critical question remains: Do these receptive field changes actually contribute to task performance? The essential computation required by either behavior is to map individual stimuli from the two sound classes (reference noise or target tone) to appropriate motor responses. It has been shown that during learning, auditory cortex undergoes long-term plasticity so that neurons explicitly encode task-relevant stimulus classes, thus facilitating the sensorimotor transformation (24, 32, 35). Similar models have been used to characterize the rapid top-down effects of attention (6, 7, 9, 10, 13). We wondered whether the changes observed here reflected an emergent representation of stimulus class, independent of the spectrotemporal details of the stimuli.

We developed a simple analytical model to predict optimal changes in a neural population for discriminating between reference and target. The excitability of A1 neurons could be selectively increased or decreased to maximize the ability of a linear decoder to identify stimulus class (SI Materials and Methods). This simulation predicted maximal discriminability by either an increase or decrease in the gain of neurons with BF near the target frequency and a compensatory change in the opposite direction for other neurons (Fig. S4). Thus, either a selective decrease or increase at the target frequency, as observed, respectively, during approach and avoidance, satisfies the discriminability constraint. This simulation highlights the fact that there are many effective ways to group target and reference sounds in the absence of other constraints. Despite their opposite sign, the changes during both tasks are optimal in this sense.

The simulation predicted that a linear decoder using the responses of A1 neurons should be able to better discriminate between reference and target classes during either behavior than during passive listening. To test this prediction, we measured the ability of a decoder using single-trial activity of subsets of neurons from one or the other behavior (Fig. 5 C and D). For both tasks, about 25% fewer neurons were required, on average, to achieve 90% accuracy during behavior than during passive listening (P < 0.001). A nonparametric analysis of mutual information, requiring fewer assumptions than the linear decoder, also revealed improved encoding of stimulus class during both behaviors (Fig. S5). Thus, although changes in spectrotemporal tuning differed between behaviors, the enhanced discriminability was consistent with emergent representation of stimulus class in A1 in both cases.

Discussion

Learning how behavior influences sensory representations requires understanding the relationship between task structure and feedback mechanisms that produce changes in sensory cortex. Depending on learned reward associations, a stimulus can trigger qualitatively different behavioral responses, such as to initiate a new behavior, cease an ongoing behavior, or make a decision between multiple choices. In this study, we manipulated stimulus–reward associations and thereby reversed the behavior (target approach or avoidance) associated with identical auditory stimuli. We measured receptive fields in A1 neurons under these different conditions and observed rapid, selective changes at target frequency of equal magnitude but opposite sign. These changes were driven by task structure, rather than by adaptation (8, 36) or peripheral motor effects. Our results highlight a unique link between rapid plasticity and behavior, and point in a new direction for deciphering the variety of task-dependent effects that have been observed in auditory and other sensory systems (2–5, 7, 8, 11, 12, 14–16, 37).

What accounts for the specific changes during the two tasks? Previously, using data only from target avoidance tasks, we proposed a model in which tuning changes served to maximize contrast between task categories (8, 13, 38), largely consistent with competitive selection models of attention (2, 6, 9, 10). The model was particularly effective in explaining plasticity in a modified task that required discrimination between two tones (27). In successive conditions where the class of a tone was reversed from reference to target, the sign of receptive field plasticity also inverted from suppression for reference tones (as with the approach data here) to enhancement for target tones (as with the avoidance data here), even within single A1 neurons. To achieve effective stimulus discrimination, the model predicted the opposite changes in reference and target responses, but the absolute sign of the changes was arbitrary. This latter constraint reflected other aspects of behavior, beyond the scope of the original model.

The current study offers insight into the possible source of top-down control. In both the approach and avoidance tasks, neural responses increased to stimuli that signaled to the animal to withhold licking. Refraining from licking reflects the inhibition of appetitive drive, a function long associated with the frontal cortex (FC) (1, 39). As we have shown previously, FC neurons respond to targets but not reference stimuli during avoidance, when targets signal behavioral inhibition (31). These A1 and FC responses are compatible with a model in which feedback to the sensory cortex reciprocally amplifies behavioral inhibition signals for downstream motor areas. Given these findings, we see that suppression of A1 reference responses during avoidance (27) may reflect the same mechanisms as the suppression of target responses during approach in our current study. These combined results permit two testable predictions. First, in A1 an approach task requiring two-tone discrimination should produce enhanced reference tone responses and suppressed target tone responses. Second, unlike the case where FC neurons responded only to targets during avoidance (31), FC neurons should respond to reference sounds during an approach task.

The dependence of STRF changes on approach vs. avoidance behavior (and, concomitantly, on stimulus reward value) suggests that top-down feedback to A1 differs between the two tasks. Although the specific modulatory networks involved in the tasks reported here are not yet known, valence-sensitive responses in the striatum (40) and habenula (41) and valence-neutral salience responses in the basal forebrain (42) encode sufficient information to differentially entrain control strategies for our tasks.

The rapid dynamics of the receptive field changes observed here clearly occur on a shorter time scale than long-term plasticity from learning (32, 35). It remains an open question how the two phenomena are related. Studies of long-term plasticity have shown that multiple modulatory systems produce plasticity (34, 43–46) and that specific effects depend on task structure (17, 20, 23, 33). However, the exact relationship between these effects and attention-driven plasticity is complex. Studies of long-term effects of classical and instrumental conditioning in A1 with either positive or negative reward have reported enhanced sensory responses at a multiunit level (47, 48) (although changes are more diverse in single units) (49). If the brain simply enhanced responses during behavior, regardless of stimulus valence, this plasticity would not facilitate discrimination between stimuli with opposite reward value. Instead, our data suggest that rapid attention-driven plasticity is bidirectional. Further studies of neuromodulatory influences during behavior can identify which aspects of rapid and long-term changes arise from common feedback networks.

Sensory behavior is influenced by numerous factors beyond stimulus class and associated motor response (e.g., sensory modality, response timing, motivation, etc.). This study focused on just two behaviors that differ along several of these dimensions, and it is not possible to point definitively to a single factor that controls the sign of rapid plasticity in A1, particularly for behaviors that differ in structure (4, 14, 16). Our findings motivate future experiments that vary behavioral factors such as motivation, reward, and task structure parametrically (20, 23, 28) to clarify the role of behavioral inhibition and other neural mechanisms in the control of sensory representations.

It has been suggested that working memory and motor planning signals for eye movements drive spatial attention effects in the visual cortex (2, 18). Just as tonotopy in the topography of the auditory system is analogous to retinotopy in the visual system, an analogous visual task to the tone vs. noise discrimination in this study is covert spatial attention to a visual scene. During these visuospatial tasks, subjects maintain fixation and inhibit saccades to the attended location, and a behavioral inhibition signal could also produce the enhanced sensory responses observed during these behaviors (5, 12, 18). Based on the findings in this study, we predict that it should be possible to suppress visual responses with an appropriate approach behavior paradigm.

More broadly, these findings suggest that the form of learning influences how information is actively processed, stored, and recalled in the brain. Developing a deeper understanding of how learned reward and motor contingencies control sensory processing may lead to more effective approaches to behavioral training and a more complete picture of how sensory information from diverse behavioral contexts is integrated into a unified representation.

Materials and Methods

Spiking activity was recorded from 501 auditory-tuned single A1 neurons of 11 awake, behaving ferrets (270 approach, 247 avoidance behavior). Animals were trained on either one behavioral paradigm (n = 8) or on both (n = 3). Both tasks required go/no-go discrimination between a pure tone target and a series of reference sounds composed of temporally orthogonal ripple combinations (26), either by licking a spout after target onset to receive a positive water reward (approach task) or ceasing to lick the spout after target offset to avoid a mild shock (avoidance task). For neurophysiological recordings, animals were implanted with a head-post and trained to perform the task while their heads were held fixed. Only data collected from experiments in which animals performed significantly better than chance were used for analysis. A subset of the avoidance data (54/247 neurons) was published previously (8). All experimental procedures conformed to standards of the National Institutes of Health and the University of Maryland Animal Care and Use Committee.

Changes in A1 Tuning During Behavior.

Tuning properties of A1 neurons were characterized by the STRF estimated by reverse correlation from responses to the reference sounds used during behavior (26). For a stimulus spectrogram, s(x,t) (frequency channel, x = 1 … X, and time, t = 1…T), and instantaneous firing rate, r(t), the STRF is the linear mapping,

Each coefficient of h indicates the gain applied to frequency x at time lag u. Positive values indicate components of the stimulus correlated with increased firing, and negative values indicate components correlated with decreased firing.

Behaviorally driven changes in tuning were determined from the difference between STRFs estimated during behavior and passive listening. Changes at the target frequency were measured as the mean difference in a window ±0.25 octaves around the target frequency at peak time lag, normalized by the peak amplitude of the passive STRF. The average STRF difference across experiments was the mean across neurons, taken after aligning each STRF difference on the frequency axis at the target frequency for that experiment. Significant changes within STRFs, across the population, and in raw neural responses were determined by a jackknifed t test.

Reference-Target Decoder.

Changes in the population representation of task stimulus class were measured by training a linear decoder to discriminate between reference and target sounds from a subset of neural responses randomly selected from the set of approach or avoidance data. Performance under different conditions (approach or avoidance, passive listening or behaving) was measured as the number of neurons required on average to achieve 90% accuracy. Significant differences in decoder performance were determined by a jackknifed t test.

Detailed methods appear in SI Materials and Methods.