Balancing the Robustness and Efficiency of Odor Representations during Learning

Abstract

For reliable stimulus identification, sensory codes have to be robust by including redundancy to combat noise, but redundancy sacrifices coding efficiency. To address how experience affects the balance between the robustness and efficiency of sensory codes, we probed odor representations in the mouse olfactory bulb during learning over a week, using longitudinal two-photon calcium imaging. When mice learned to discriminate between two dissimilar odorants, responses of mitral cell ensembles to the two odorants gradually became less discrete, increasing the efficiency. In contrast, when mice learned to discriminate between two very similar odorants, the initially overlapping representations of the two odorants became progressively decorrelated, enhancing the robustness. Qualitatively similar changes were observed when the same odorants were experienced passively, a condition that would induce implicit perceptual learning. These results suggest that experience adjusts odor representations to balance the robustness and efficiency depending on the similarity of the experienced odorants.

Figure 1. Longitudinal mitral cell imaging during week-long behavioral paradigms

(A) Schematic demonstrating the tradeoff between robustness and efficiency in the encoding of stimuli within a finite neural activity space (rectangles). (top) An extreme example of robust coding: Three discrete sensory stimuli (colored circles) are encoded with high robustness, or redundancy. (bottom) An example of a system with higher efficiency, or the capacity to encode more stimuli, than the system above within the same neural space. However, this enhanced efficiency occurs at the expense of robustness. (B) Schematic of the olfactory bulb. AAV2.1-FLEX-hsyn-GCaMP6f was injected into the right olfactory bulb of Pcdh21-cre mice to express GCaMP6f specifically in mitral cells. (C) Experimental timeline. Mice first go through a pre-training period with two sets of odorant pairs (first pre-training pair: citral/limonene; second pre-training pair: +-carvone/cumene) before they started the imaging period, where they perform the discrimination task with a novel odorant pair. (D) Trial structure of the discrimination task. (E) Schematic of imaging setup. (F) A field of mitral cells expressing GCaMP6f on the first day of imaging (left) and six days later (right).

Figure 2. Mitral cell odorant responses during the easy discrimination task

(A) Left: Behavioral performance on Day 1 of the easy discrimination task. Fraction of correct trials is shown for each block of 10 trials (n = 8 mice). Right: Behavioral performance for each session (day) during the easy discrimination task. (B) Mean odorant responses of three example mitral cells during single sessions. Horizontal bars indicate odorant periods (4 sec). (C) Spatial distribution of responsive (cyan) and divergent (magenta) neurons on Day 1 (left) and Day 7 (right) for an example mouse during easy discrimination training. Non-responsive neurons are shown in white. (D) Fractions of neurons classified as responsive (black) and divergent (magenta) on each day. Both fractions show a significant decrease (Pearson correlation; Responsive: r = -0.43, p < 0.01, divergent: r = - 0.44, p < 0.01). (E) Fraction of divergent neurons out of responsive neurons is maintained throughout easy discriminating training (Pearson correlation; r = -0.03, p = 0.82). (F) Sensitivity index (d′) of divergent neurons decreases with easy discrimination training (Pearson correlation; r = -0.28, p < 0.05). (G) Decoder accuracy during easy discrimination training significantly decreases (Pearson correlation; r = -0.42, p < 0.01) (H) Coding of odorant identity in mitral cell ensembles is distributed. For each mouse, the mean decoder accuracy was calculated after removing one additional neuron at a time in the descending order of their contribution to decoder accuracy (i.e. the drop in decoder accuracy caused by removal, Methods). (All values in line plots are mean ± SEM.)

Figure 3. Mitral cell odorant responses during the difficult discrimination task

(A) Behavioral performance during the difficult discrimination task (n = 10 mice). (B) Mean odorant responses of three example mitral cells during difficult discrimination training. Horizontal bars indicate odorant periods (4 sec). (C) Fractions of neurons classified as responsive (black) and divergent (magenta) on each day. The responsive fraction shows a significant decrease (Pearson correlation; r = -0.45, p < 0.001), while the divergent fraction remains stable (Pearson correlation; r = 0.20, p = 0.10). (D) Spatial distribution of responsive (cyan) and divergent (magenta) neurons on Day 1 (left) and Day 7 (right) for an example mouse during difficult discrimination training. (E) Fraction of divergent neurons out of responsive neurons increases throughout difficult discriminating training (Pearson correlation; r = 0.41, p < 0.001). (F) The sensitivity index (d′) of divergent neurons increases with difficult discrimination training (Pearson correlation; r = 0.31, p < 0.05). (G) Population decoder accuracy is enhanced during difficult discrimination training (Pearson correlation; r = 0.38, p < 0.01). (H) Coding of odorant identity in mitral cell ensembles is distributed. (I) Improvement in decoder accuracy is not due to changes in interneuronal correlation structure. When each odorant's trial responses were shuffled independently for each cell to eliminate noise correlation, decoder accuracy slightly improved, and this effect did not change with training (Pearson correlation; r = 0.23, p = 0.07). (J) The trial-by-trial variance of ensemble activity, calculated as the sum of the covariance matrix of a session's activity vector (Methods), does not change with learning (Pearson correlation; r = -0.04, p = 0.72). (K) Mean square distance between odor centroids (Methods) significantly increases with learning (Pearson correlation; r = 0.35, p < 0.01). (L) Mean square distance between odor centroids and the decoder accuracy are highly correlated (Pearson correlation; r = 0.79, p < 0.001). (M) Distribution of all hit, correct rejection (CR), and false alarm (FA) responses on Day 7, projected on the axis connecting the two odor centroids. CR and FA distributions are distinct (bootstrap, p < 0.001) and the FA trial distribution lies between hit and CR trials. (N) Probability of false alarms in all odorant 2 trials is higher when mitral cell responses are more similar to odorant 1 trials on Day 7 (Spearman correlation; r = -0.69, p < 0.001). (All values in line plots are mean ± SEM.)

Figure 4. Bidirectional changes in the divergence of population representations during the easy and difficult discrimination tasks

(A) Mitral cell population responses from a single mouse in the easy discrimination task visualized in the space of the first three principal components (PC) on Day 1 (top) and Day 7 (bottom). Each data point corresponds to the population activity on a trial. (B) Mitral cell population responses pooled across all animals in the easy discrimination task, plotted for Day 1 and Day 7 in the first three PC axes. Note the decrease in separation of odorant 1 and odorant 2 trials with training. (C) and (D) Same as A and B for the difficult discrimination task. Note the increase in separation of odorant 1 and odorant 2 trials with training. The value next to each PC axis label is the variance accounted for by that PC axis, and plots are manually rotated to optimally highlight any separability between odorants.

Figure 5. Mitral cell odor responses during passive experience

(A-C): Passive experience of the same odorants used in the easy discrimination task (n = 8 mice): (A) Fraction of neurons classified as responsive (black) and divergent (magenta) on each day. There is a significant decrease of both fractions (Pearson correlation; responsive: r = -0.38, p < 0.01; divergent: r = - 0.37, p < 0.01). (B) Fraction of divergent neurons out of responsive neurons is stable during the week- long passive exposure (Pearson correlation; r = 0.05, p = 0.69). (C) Decoder accuracy significantly decreases (Pearson correlation; r = -0.29, p < 0.05). (D-F): Same as A-C for passive experience of the same odorants used in the difficult discrimination task (n = 11 mice): (D) Fraction of neurons classified as responsive decreases (Pearson correlation; r = -0.42, p < 0.001), while the divergent fraction does not change (Pearson correlation, r = 0.06, p = 0.61). (E) Fraction of divergent neurons out of responsive neurons increases (Pearson correlation; r = 0.35, p < 0.01). (F) Decoder accuracy is stable (Pearson correlation, r = 0.18, p = 0.11). (All values in line plots are mean ± SEM.)