Dynamic contrast enhancement and flexible odor codes

Sensory stimuli evoke spiking activities patterned across neurons and time that are hypothesized to encode information about their identity. Since the same stimulus can be encountered in a multitude of ways, how stable or flexible are these stimulus-evoked responses? Here we examine this issue in the locust olfactory system. In the antennal lobe, we find that both spatial and temporal features of odor-evoked responses vary in a stimulus-history dependent manner. The response variations are not random, but allow the antennal lobe circuit to enhance the uniqueness of the current stimulus. Nevertheless, information about the odorant identity is conf ounded due to this contrast enhancement computation. Notably, predictions from a linear logical classifier (OR-of-ANDs) that can decode information distributed in flexible subsets of neurons match results from behavioral experiments. In sum, our results suggest that a trade-off between stability and flexibility in sensory coding can be achieved using a simple computational logic.

method with leave-one-out cross-validation in 85-D space was used for generating these classification results (see Methods). Note that the confusion matrix is mostly diagonal indicating that the PN responses evoked by the same stimulus presented with different histories are distinct.
(d) Similar plot as in panel c but the confusion matrix analyzing the response separability of the solitary and sequential geraniol presentations is shown.
Supplementary Figure 2 (a) Similar trajectory plots as shown in Fig. 2a,b but trial-to-trial variations are included. Population PN responses evoked by the distractor odorant, hex, and the sequential presentation of hex are shown for three sets of trials: mean of trials 1-3, mean of trials 4-6, and mean of trials 7-10. The darker colored traces correspond to the earlier set of trials.
(b) Similar plots as in panel a but shown for five sequential presentations of geraniol.
(c, d) Similar plots as in Fig. 2c,d, but the correlations are now computed using the mean odorevoked response during the last 1 s of stimulus presentation window. Figure 3 (a) Comparison between the mean firing rates averaged across two distinct sets of PNs is shown for the solitary (red trace) and sequential presentations (orange traces) of hex. Overlapping PNs correspond to the set of PNs that were responsive to both the target (hex) and distractor odorant. 'Non-overlapping PNs' correspond to the remaining set of PNs that were not 'overlapping PNs'. Left panel shows the firing rate of overlapping PNs averaged across trials (n = 10). Right panel shows the firing rate of non-overlapping PNs averaged across ten trials. The percentage of overlapping PNs (i.e. co-activation) for each distractor odorant is shown.   (a-f) Palp-opening responses (POR) to additional distractor-target odor sequences are shown. The distance between the palps was tracked and plotted as a function of time. Error bar represents standard error across locusts (n = 20).

Supplementary Figure 7
Classification results for the digital version of the flexible set decoder are shown. For any 50 ms time bin, the threshold for ON classification was set to be 3-of-36 PNs must be responding to the stimulus presentation (i.e. firing rate > 6.5 s.d. of pre-stimulus activity). The threshold for OFF classification was also set to be 3-of-33 PNs. The 36 hex-ON PNs and 33 hex-OFF PNs were determined based on solitary hexanol presentations alone. Note that these classification results are very similar to the analog version that we presented in Fig. 7b.  (a, b) A schematic illustration of the flexible set decoder is shown. Here the problem is cast as one of object recognition (i.e. chair recognition). Images of different chairs and their features are tabulated. Photo courtesy of Raman Lab.
(c) A schematic of an OR-of-ANDs or disjunction-of-conjunction classifier is shown. Input features are binary 'feature present' or 'feature absent'. The weight vectors are constant and set based on an ideal object (Chair 2 in panel a): the weight vector component is a '1' if the feature is present in the ideal chair, and '0' if it not present. The only free parameter in the OR-of-ANDs classifier is the threshold of the output node (q). If the value of the threshold is set below the total number of features present in the ideal object (for example q = 2), then the presence of any two of the four features will allow recognition of the object (i.e. flexible decoding). This can be written as a set of logical OR-of-ANDs operation.
(d) The list of analogies between this object/chair recognition illustration and flexible odor decoding proposed in this manuscript are listed in a table.
Supplementary Figure 10 (a, b) Response trajectories generated by two additional sequential presentations of hex are plotted: 2oct -2 s -hex and 2oct -10 s -hex. Similar method as in Fig. 2a was followed to analyze this dataset.
(c) The mean of correlation values between the ensemble PN responses (n = 104 PNs) evoked by hexanol and the three distractor cues (2oct, iaa, and chex) are shown as bar plots. Error bars indicate ± s.d across ten trials. The mean odor-evoked responses during the initial 1 s after stimulus onset were used for computing these correlations. The odor-pairs that were compared are identified along the x-axis. Asterisks indicate a significant decrease in the correlation (*P < 0.025 (Bonferroni corrected for two comparisons), t-tests, n = 10 trials). hex(solitary) stimulus. To allow for a fair comparison, the correlations for inter-condition correlations (plotted in orange) were calculated by estimating the similarity between the mean PN response in the first five trials of hex(solitary) exposures with the mean response in the first five trials of sequential hex presentations. This was again done for each PN and for each sequential hex presentation to generate the five orange distributions shown in the plot.
(b) Similar plots as in panel a but analyzing PN responses to geraniol.
(c) Comparison of combinatorial PN response profiles activated by the same odorant across trials, same odorant across stimulus histories, and between different odorants is shown as a function of trial number. Note that all comparisons are made with respect to the ensemble PN responses elicited by solitary presentation of hexanol in the very first trial.
(d) Similar plots as panel c but plotted when the target odor is ger.