Abstract
Affective learning and memory are essential for daily behavior, with both adaptive and maladaptive learning depending on stimulusevoked activity in the amygdala circuitry. Behavioral studies further suggest that postassociation offline processing contributes to memory formation. Here we investigated spike sequences across simultaneously recorded neurons while monkeys learned to discriminate between aversive and pleasant tone–odor associations. We show that triplets of neurons exhibit consistent temporal sequences of spiking activity that differed from firing patterns of individual neurons and pairwise correlations. These sequences occurred throughout the long posttrial period, contained valencerelated information, declined as learning progressed and were selectively present in activity evoked by the recent pairing of a conditioned stimulus with an unconditioned stimulus. Our findings reveal that temporal sequences across neurons in the primate amygdala serve as a coding mechanism and might aid memory formation through the rehearsal of the recently experienced association.
Access options
Subscribe to Journal
Get full journal access for 1 year
$225.00
only $18.75 per issue
All prices are NET prices.
VAT will be added later in the checkout.
Rent or Buy article
Get time limited or full article access on ReadCube.
from$8.99
All prices are NET prices.
Data availability
All data supporting the findings of this study are available from the corresponding author upon reasonable request.
Code availability
Custom code for behavioral and electrophysiological tests is available from the corresponding author upon reasonable request.
References
 1.
Baxter, M. G. & Murray, E. A. The amygdala and reward. Nat. Rev. Neurosci. 3, 563–573 (2002).
 2.
Janak, P. H. & Tye, K. M. From circuits to behaviour in the amygdala. Nature 517, 284–292 (2015).
 3.
Salzman, C. D., Paton, J. J., Belova, M. A. & Morrison, S. E. Flexible neural representations of value in the primate brain. Ann. NY Acad. Sci. 1121, 336–354 (2007).
 4.
SugaseMiyamoto, Y. & Richmond, B. J. Neuronal signals in the monkey basolateral amygdala during reward schedules. J. Neurosci. 25, 11071–11083 (2005).
 5.
Herry, C. & Johansen, J. P. Encoding of fear learning and memory in distributed neuronal circuits. Nat. Neurosci. 17, 1644–1654 (2014).
 6.
Maren, S. & Quirk, G. J. Neuronal signalling of fear memory. Nat. Rev. Neurosci. 5, 844–852 (2004).
 7.
Krabbe, S., Grundemann, J. & Luthi, A. Amygdala inhibitory circuits regulate associative fear conditioning. Biol. Psychiatry 83, 800–809 (2018).
 8.
Duvarci, S. & Pare, D. Amygdala microcircuits controlling learned fear. Neuron 82, 966–980 (2014).
 9.
Namburi, P. et al. A circuit mechanism for differentiating positive and negative associations. Nature 520, 675–678 (2015).
 10.
Yu, K. et al. The central amygdala controls learning in the lateral amygdala. Nat. Neurosci. 20, 1680–1685 (2017).
 11.
Josselyn, S. A., Kohler, S. & Frankland, P. W. Finding the engram. Nat. Rev. Neurosci. 16, 521–534 (2015).
 12.
Likhtik, E. & Paz, R. Amygdala–prefrontal interactions in (mal)adaptive learning. Trends Neurosci. 38, 158–166 (2015).
 13.
Averbeck, B. B. & Chafee, M. V. Using model systems to understand errant plasticity mechanisms in psychiatric disorders. Nat. Neurosci. 19, 1418–1425 (2016).
 14.
Delgado, M. R., Olsson, A. & Phelps, E. A. Extending animal models of fear conditioning to humans. Biol. Psychol. 73, 39–48 (2006).
 15.
Milad, M. R. & Quirk, G. J. Fear extinction as a model for translational neuroscience: ten years of progress. Annu. Rev. Psychol. 63, 129–151 (2012).
 16.
Johansen, J. P. et al. Optical activation of lateral amygdala pyramidal cells instructs associative fear learning. Proc. Natl Acad. Sci. USA 107, 12692–12697 (2010).
 17.
Quirk, G. J., Repa, C. & LeDoux, J. E. Fear conditioning enhances shortlatency auditory responses of lateral amygdala neurons: parallel recordings in the freely behaving rat. Neuron 15, 1029–1039 (1995).
 18.
Herry, C. et al. Switching on and off fear by distinct neuronal circuits. Nature 454, 600–606 (2008).
 19.
Johansen, J. P., Cain, C. K., Ostroff, L. E. & LeDoux, J. E. Molecular mechanisms of fear learning and memory. Cell 147, 509–524 (2011).
 20.
Pape, H. C. & Pare, D. Plastic synaptic networks of the amygdala for the acquisition, expression, and extinction of conditioned fear. Physiol. Rev. 90, 419–463 (2010).
 21.
Sah, P., Westbrook, R. F. & Luthi, A. Fear conditioning and longterm potentiation in the amygdala: what really is the connection? Ann. NY Acad. Sci. 1129, 88–95 (2008).
 22.
Gibbon, J., Baldock, M. D., Locurto, C., Gold, L. & Terrace, H. S. Trial and intertrial durations in autoshaping. J. Exp. Psychol. Anim. Behav. Process. 3, 264–284 (1977).
 23.
Lattal, K. M. Trial and intertrial durations in Pavlovian conditioning: issues of learning and performance. J. Exp. Psychol. Anim. Behav. Process. 25, 433–450 (1999).
 24.
Wagner, A. R., Rudy, J. W. & Whitlow, J. W. Rehearsal in animal conditioning. J. Exp. Psychol. 97, 407–426 (1973).
 25.
Wagner, A. R. SOP: a model of automatic memory processing. in Information Processing in Animals: Memory Mechanisms (eds Spear, N. E. & Miller, R. R.) 5–47 (Psychology Press, 1981).
 26.
Nader, K., Schafe, G. E. & LeDoux, J. E. The labile nature of consolidation theory. Nat. Rev. Neurosci. 1, 216–219 (2000).
 27.
Holland, P. C. & Schiffino, F. L. Minireview: prediction errors, attention and associative learning. Neurobiol. Learn. Mem. 131, 207–215 (2016).
 28.
McIntyre, C. K., Power, A. E., Roozendaal, B. & McGaugh, J. L. Role of the basolateral amygdala in memory consolidation. Ann. NY Acad. Sci. 985, 273–293 (2003).
 29.
Lee, S. C., Amir, A., Haufler, D. & Pare, D. Differential recruitment of competing valencerelated amygdala networks during anxiety. Neuron 96, 81–88.e5 (2017).
 30.
Belova, M. A., Paton, J. J. & Salzman, C. D. Momenttomoment tracking of state value in the amygdala. J. Neurosci. 28, 10023–10030 (2008).
 31.
Pryluk, R., Kfir, Y., GelbardSagiv, H., Fried, I. & Paz, R. A tradeoff in the neural code across regions and species. Cell 176, 597–609.e18 (2019).
 32.
Schneidman, E., Berry, M. J. 2nd, Segev, R. & Bialek, W. Weak pairwise correlations imply strongly correlated network states in a neural population. Nature 440, 1007–1012 (2006).
 33.
Martignon, L. et al. Neural coding: higherorder temporal patterns in the neurostatistics of cell assemblies. Neural Comput. 12, 2621–2653 (2000).
 34.
Nakahara, H. & Amari, S. Informationgeometric measure for neural spikes. Neural Comput. 14, 2269–2316 (2002).
 35.
Panzeri, S. & Schultz, S. R. A unified approach to the study of temporal, correlational, and rate coding. Neural Comput. 13, 1311–1349 (2001).
 36.
Girardeau, G., Inema, I. & Buzsaki, G. Reactivations of emotional memory in the hippocampus–amygdala system during sleep. Nat. Neurosci. 20, 1634–1642 (2017).
 37.
Feldman, D. E. The spiketiming dependence of plasticity. Neuron 75, 556–571 (2012).
 38.
Carr, M. F., Jadhav, S. P. & Frank, L. M. Hippocampal replay in the awake state: a potential substrate for memory consolidation and retrieval. Nat. Neurosci. 14, 147–153 (2011).
 39.
Buzsaki, G. & Llinas, R. Space and time in the brain. Science 358, 482–485 (2017).
 40.
O’Keefe, J. & Dostrovsky, J. The hippocampus as a spatial map. Preliminary evidence from unit activity in the freelymoving rat. Brain Res. 34, 171–175 (1971).
 41.
Schnitzer, M. J. & Meister, M. Multineuronal firing patterns in the signal from eye to brain. Neuron 37, 499–511 (2003).
 42.
Ikegaya, Y. et al. Synfire chains and cortical songs: temporal modules of cortical activity. Science 304, 559–564 (2004).
 43.
Pillow, J. W. et al. Spatiotemporal correlations and visual signalling in a complete neuronal population. Nature 454, 995–999 (2008).
 44.
Ganmor, E., Segev, R. & Schneidman, E. Sparse loworder interaction network underlies a highly correlated and learnable neural population code. Proc. Natl Acad. Sci. USA 108, 9679–9684 (2011).
 45.
Oram, M. W., Wiener, M. C., Lestienne, R. & Richmond, B. J. Stochastic nature of precisely timed spike patterns in visual system neuronal responses. J. Neurophysiol. 81, 3021–3033 (1999).
 46.
Buzsaki, G. Neural syntax: cell assemblies, synapsembles, and readers. Neuron 68, 362–385 (2010).
 47.
Grewe, B. F. et al. Neural ensemble dynamics underlying a longterm associative memory. Nature 543, 670–675 (2017).
 48.
Reijmers, L. G., Perkins, B. L., Matsuo, N. & Mayford, M. Localization of a stable neural correlate of associative memory. Science 317, 1230–1233 (2007).
 49.
Rashid, A. J. et al. Competition between engrams influences fear memory formation and recall. Science 353, 383–387 (2016).
 50.
Grundemann, J. & Luthi, A. Ensemble coding in amygdala circuits for associative learning. Curr. Opin. Neurobiol. 35, 200–206 (2015).
 51.
Livneh, U. & Paz, R. Amygdala–prefrontal synchronization underlies resistance to extinction of aversive memories. Neuron 75, 133–142 (2012).
 52.
Livneh, U. & Paz, R. Aversivebias and stageselectivity in neurons of the primate amygdala during acquisition, extinction, and overnight retention. J. Neurosci. 32, 8598–8610 (2012).
 53.
Livneh, U. & Paz, R. An implicit measure of olfactory performance for nonhuman primates reveals aversive and pleasant odor conditioning. J. Neurosci. Methods 192, 90–95 (2010).
 54.
Harrison, M. T., Amarasingham, A. & Truccolo, W. Spatiotemporal conditional inference and hypothesis tests for neural ensemble spiking precision. Neural Comput. 27, 104–150 (2015).
 55.
Nauhaus, I., Busse, L., Carandini, M. & Ringach, D. L. Stimulus contrast modulates functional connectivity in visual cortex. Nat. Neurosci. 12, 70–76 (2009).
 56.
Panzeri, S., Senatore, R., Montemurro, M. A. & Petersen, R. S. Correcting for the sampling bias problem in spike train information measures. J. Neurophysiol. 98, 1064–1072 (2007).
 57.
Maoz, O. & Schneidman, E. maxent_toolbox: Maximum entropy toolbox for MATLAB, version 1.0.2. 2017. https://orimaoz.github.io/maxent_toolbox (2017).
 58.
Joshua, M., Elias, S., Levine, O. & Bergman, H. Quantifying the isolation quality of extracellularly recorded action potentials. J. Neurosci. Methods 163, 267–282 (2007).
Acknowledgements
The authors thank Y. Kfir, A. Taub, U. Livneh, Y. Cohen, K. Aberg, E. Schneidman and E. Karpas for scientific consultations. The authors thank Y. Shohat for animal training, experiments and welfare, E. Kahana for medical and surgical procedures, and E. FurmanHaran and F. Attar for MRI procedures. This work was supported by ISF no. 2352/19 and ERC2016CoG no. 724910 grants to R.P.
Author information
Affiliations
Contributions
T.R.S. and R.P. conceived and designed the experiments. T.R.S. planned and performed the analyses. T.R.S. and R.P. wrote the manuscript.
Corresponding author
Correspondence to Rony Paz.
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Extended data
Extended data Fig. 1 Structure in triplets (sequences of threespikes from three neurons).
(A) Circular shuffling illustration. Shuffling was performed on 10s activity segments. Activity of two randomly chosen units (to control for single neuron activity) was circularly shifted by a random lag between 150–300 ms, one left and the other right. The spiking structure of each neuron, and firing rate correlations beyond 300–600 ms, are preserved, yet the precise spike time correlation between the neurons is destroyed. (B) Structure analysis illustration. Right top: Triplet 177 sequence probability distribution (log scale). Right bottom: Examples of triplet 177 shuffled data sequence probability distributions (two shuffled units). Left: data distribution (front) and 30 examples of shuffled data distributions. Structure was examined by calculating (1) the mean JensenShannondivergence (JSD) dissimilarity between sequences of the data and each of the shuffled data (\(\bar D_{data}\), blue arrow, comparing right top and left distributions), and (2) the mean JSD dissimilarity between sequences of each shuffled data instance and all other instances (\(\bar D_{shuffle}^{\left\{ {1, \ldots n} \right\}}\), green arrow, comparing right bottom and left distributions). If the dissimilarity between the data and the shuffled data is larger than the dissimilarities between the shuffled data sets, it suggests that the structure of the sequence distribution is different than expected from single neuron activity. Shaded area: s.e.m.
Extended data Fig. 2 Locality of sequence activity.
(A) Average maximal physical distance between the three electrodes (taking only triplets recorded from three different electrodes) for triplets exhibiting sequence structure (n=24) and triplets with no structure (n=80). Dots represent maximal distance between electrodes of a single triplet. Structured triplets were physically closer than nonstructured triplets (one tailed Wilcoxon rank sum: Z=−2.47, p=0.007, r_{rb}=0.24). (B) Average maximal physical distance between the three electrodes (taking only triplets recorded from three different electrodes) for consistent triplets (n=16) and triplets with no significant consistency (n=88). Dots represent maximal distance between electrodes of a single triplet. Consistent triplets were physically closer than nonconsistent triplets (one tailed Wilcoxon rank sum test: Z=−3.1, p < 0.001, r_{rb} = 0.3). Error bars: s.e.m. **p<0.01.
Extended data Fig. 3 Consistency in sequences.
Top: Triplet 177 estimated sequence probability distribution in two subdivisions of the time period (left and right, sequence order is as in Fig. 2a left). Bottom: Shuffled data sequence probability distributions of the same subdivisions. Consistency was examined by calculating (1) the mean JSD dissimilarity between sequences of the data in the two subdivisions, for 100 subdivisions (\(\bar C_{data}\), blue arrow, comparing left and right top panels), and (2) the mean JSD dissimilarity between sequences of the data of one half and all shuffled instance of the other half (\(\bar C_{shuffle}^{\left\{ {1, \ldots n} \right\}}\), green arrows, comparing (a) left top and right bottom panels and (b) right top and left bottom panels), for 100 subdivisions. If the dissimilarity between the data subdivisions is smaller than the dissimilarities between the data and the shuffled data sets, it suggests that the structure of sequences is consistent beyond expected from single neuron activity. Shaded area: s.e.m.
Extended data Fig. 4 Comparing Amygdala to ACC sequences.
(A) proportion of significantly structured (BenjaminiHochberg [BH] corrected for multiple comparisons) amygdala (blue, n=355) and dACC (green, n=564) triplets. The proportion was higher in amygdala than in dACC triplets (χ^{2} test for independence:\({\mathrm{\chi }}_{{\mathrm{df}} = 1}^2 = 20.5,\,\mathrm{p} < 10^{\mathrm{5}}\)). (B) Violin plots of the structure scores of individual triplets (black dots, only significant triplets). The mean score (black dashed line) of amygdala triplets (blue, n=173) was higher than that of dACC (green, n=189, independent twotailed ttest, t_{360} = 4.9, p < 10^{−5}, d=0.51). The colored surface marks the kernel density estimate of the corresponding probability distribution, the thick gray line marks the interquartile range and the white dot marks the median. (CD) Consistency analysis—arranged as A,B. Higher proportion in amygdala (n=355) than in dACC (n=564): \({\mathrm{\chi }}_{\mathrm{df} = 1}^2 = 30.7,\,{\mathrm{p}} < 10^{  7}\); Higher score in amygdala (n=116) than in dACC (n=94): t_{208}=5.9, p < 10^{−7}, d=0.83. Error bars: s.e.m. ***p<0.001.
Extended data Fig. 5 Decoding performance and controls.
(A) The proportion of triplets with significant decoding hit rate (BH corrected, false discovery rate [FDR]≤0.05, blue line) was significantly higher than that of trialshuffle control (n=250 trial shuffle data set, MC pvalue: p=0.024). Inset: triplets from different recording electrodes (p=0.11). (B) The mean decoding hit rate (n=355, 0.56±0.01) across all triplets (blue line) was significantly higher than that of independent trialshuffle control (turquoise histogram, n=250 trialshuffle data set, MC pvalue: p=0.004). Inset: triplets from different recording electrodes (p=0.004). Note that x axes begin at chance level (0.5). (C) Decoding hit rate for decoding pleasant trials from pretask activity (purple) and decoding aversive trials from pretask activity (red) for all triplets that were FDR significant for decoding pleasant from aversive trials (Fig. 3a, n=101). The hit rate for simultaneous recorded activity (data, xaxis) was higher than the hit rate for independent shuffle trial control (shuffle trials, yaxis) for both aversive (one tailed ttest: t_{70}=1.71, p=0.046, d=0.13) and pleasant decoding (t_{70}=3.12, p=0.001, d=0.19). Left top inset: mean hit rate for the data (blue) and trial shuffled data (turquoise). Note that the y axis begins at chance level (0.5). Right bottom inset: Histogram of the difference between data hit rate and trial shuffle hit rate in individual triplets. For both pleasant (purple) and aversive (red) trials, the majority of triplets showed higher hit rate for the data, so that the difference is larger than 0 (dashed black line). (D) Histograms of decoding hit rates of higherthanchance triplets for sequences (blue, N=76) and for ISI (green, N=130, one tailed independent ttest: t_{204}=1.69, p=0.046, d=0.24). Note that the axes begin at chance level (0.5). In all panels error bars mark s.e.m, *p<0.05;**p<0.01.
Extended data Fig. 6 Maximum entropy (ME) models.
(A) SpatialME model fitted to quadruplets of neurons. Spike trains were binned to 50ms nonoverlapping bins (top), creating four letters spatial words (right bottom). The full model was trained on spatial correlations and the coefficients of independent rates (h_{i}), pairwise correlations (j_{ij}) and triplewise correlations (m_{ijk}) were learned. This model captures a lower dimension of sequence activity, where the temporal proximity of spikes from different neurons is preserved yet the order of spikes is lost due to binning. (B) SequenceME model: Spatiotemporal ME model fitted to sequence activity matrices in triplets. Bins with no spiking activity were removed and spike trains were kept at 1ms resolution (top). The full model was trained on two steps temporal activity of three neurons, creating a 3X3 (9 letters) spatiotemporal word (right bottom). The coefficients of independent rates (h_{i}), spatial pairwise correlations (h_{ij}), pairwise spatiotemporal correlations \(\left( {j_{i\left( t \right)j\left( {t + 1} \right)},j_{i\left( t \right)j\left( {t + 2} \right)}} \right)\) and triplewise spatiotemporal correlations \(m_{i\left( t \right)j\left( {t + 1} \right)k\left( {t + 2} \right)}\) were learned. This model captures the order of spikes between triplets of neurons, neglecting time lags between spikes and is similar to the sequence activity defined originally (as in Fig. 1c).
Extended data Fig. 7 Decoding CSUS activity using posttrial activity with ME models.
(A) Hit rate of spatialME decoding trained on the posttrial activity and tested on CSUS activity (CSUS hit rate), taking all quadruplets with betterthanchance decoding for the posttrial activity in either model (Fig. 5e, p<0.05, n=198). Main panels: violin plot of the normalized (Zscore) CSUS hit rate of individual quadruplets (black dots) for each model. The colored surface marks the kernel density estimate of the corresponding probability distribution, the thick gray line marks the interquartile range and the dashed black line marks the mean of the normalized CSUS hit rate. The pairwise and triplewise models (middle and right) showed higher rates than that of the independent (one way repeated measures ANOVA, \(F_{194}^2 = 7.9,\,p = 0.001\), one tailed Tukey HSD, independentpairwise: \(p < 10^{  3},\,d = 0.63\), independenttriplewise: \(p = 0.02,\,d = 0.49\)), as apparent from the higher means and larger densities in the top part of the two right distributions compared to the left. Left inset: proportion of highest hit rate in quadruplets using decoding based on the independent, pairwise and triplewise model. Right insets: CSUS hit rate (mean±s.e.m). Note that the y axis begins at chance level (0.5). *p<0.05;**p<0.01; n.s p>0.05. (B) Same as (A) for the sequenceME, taking triplets with betterthanchance decoding for the posttrial activity in either model (Fig. 5f, p<0.05, n=129). CSUS activity was averaged over 15 trials to ensure sufficient sampling. The triplewise models (right) showed higher rates than that of the independent (left, \(F_{194}^2 = 3.68,\,p = 0.03\), one tailed Tukey HSD, p=0.03, d=0.44) and pairwise (middle, one tailed ttest t_{89}=1.83, p=0.035, d=0.25), apparent from the higher density in the lower rates of the independent and the higher density in the higher rates of the triplewise model. Violin elements are as in A.
Extended data Fig. 8 Negligible effect of nonstationary firing rates (FR).
Stationary and nonstationary units were separated by two methods: 1. Two tailed ttest comparing the average firing rate in the first and last 150 seconds of the pretask activity (FR ttest). 2. Runstest examining if interspikeintervals along the pretask activity were drawn randomly from a single distribution. Units with significant results (p<0.05) were classified as nonstationary and all other units were classified as stationary. Only triplets with three stationary units were classified as stationary triplets (FR ttest: n=156; runstest: n=38). (A) There was no significant difference between the proportion of structured triplets in stationary (pink) and nonstationary (blue) triplets as classified by either the FR ttest (left bars, χ^{2}=2.91, p=0.09) or the runstest (right bars, χ^{2}=0.75, p=0.39), suggesting a negligible effect of stationarity on structure. (B) There was no significant difference between the proportion of consistent triplets in stationary (pink) and nonstationary (blue) triplets as classified by either the FR ttest (left bars, χ^{2}=3.35, p=0.07) or the runstest (right bars, χ^{2}=0.27, p=0.6), suggesting a negligible effect of stationarity on consistency. (C) Violin plots of the proportion of reduction in entropy by the thirdorder correlation of the sequenceME model, normalized by the proportion in pairwise surrogate data (see Fig. 5d and methods). The colored surface marks the kernel density estimate of the corresponding probability distribution, the thick gray line marks the interquartile range and the white dot marks the median. The proportion of reduction in entropy in stationary triplets (pink) was significantly larger than that of nonstationary triplets (blue) as classified by the FR ttest (left violins, two tailed rank sum test, Z=2.36, p=0.02) and there was no difference between these groups as classified by the runstest (right violins, Z=−1.03, p=0.3), suggesting a negligible (if at all) effect of stationarity on sequenceME results. (D, E) Reduction in entropy due to the triplewise correlations of the SequenceME model (see main Fig. 5d) in stationary triplets. (D) FR ttest, n=156; (E) runs test, n=38. The proportion of reduction in entropy due to the triplewise correlations (I_{3}/I_{N}) in the real data (xaxis) was larger than in surrogate data sampled from the pairwise ME distribution (pairwisesurrogate control, yaxis). Black dashed line: identity line. Inset: means and SEM over all stationary triplet. Paired t test between medians across trials; FR ttest for stationarity: t_{121}=18.08, p<10^{−30}, d=1.05; runs test: t_{20}=6.11, p<10^{−4}, d=0.68. In all panels, error bars and shaded area mark s.e.m., *p<0.05; ***p<0.001; n.s p>0.05.
Supplementary information
Supplementary Information
Supplementary Figs. 1–13.
Supplementary software 1
Custom Matlab code used for analyses.
Rights and permissions
About this article
Cite this article
ReitichStolero, T., Paz, R. Affective memory rehearsal with temporal sequences in amygdala neurons. Nat Neurosci 22, 2050–2059 (2019) doi:10.1038/s4159301905429
Received
Accepted
Published
Issue Date
DOI
Further reading

Pavlovian patterns in the amygdala
Nature Neuroscience (2019)