Skip to main content

Thank you for visiting You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

Neural ensemble dynamics underlying a long-term associative memory


The brain’s ability to associate different stimuli is vital for long-term memory, but how neural ensembles encode associative memories is unknown. Here we studied how cell ensembles in the basal and lateral amygdala encode associations between conditioned and unconditioned stimuli (CS and US, respectively). Using a miniature fluorescence microscope, we tracked the Ca2+ dynamics of ensembles of amygdalar neurons during fear learning and extinction over 6 days in behaving mice. Fear conditioning induced both up- and down-regulation of individual cells’ CS-evoked responses. This bi-directional plasticity mainly occurred after conditioning, and reshaped the neural ensemble representation of the CS to become more similar to the US representation. During extinction training with repetitive CS presentations, the CS representation became more distinctive without reverting to its original form. Throughout the experiments, the strength of the ensemble-encoded CS–US association predicted the level of behavioural conditioning in each mouse. These findings support a supervised learning model in which activation of the US representation guides the transformation of the CS representation.

This is a preview of subscription content, access via your institution

Relevant articles

Open Access articles citing this article.

Access options

Rent or buy this article

Get just this article for as long as you need it


Prices may be subject to local taxes which are calculated during checkout

Figure 1: Ca2+ imaging of BLA neural activity across a 6-day fear-conditioning protocol.
Figure 2: Fear conditioning induces bi-directional changes in BLA signalling.
Figure 3: Learning increases the similarity of the CS+ and US representations.
Figure 4: During behavioural extinction, the CS+ representation becomes more distinguishable from the US representation but does not revert to its initial form.
Figure 5: The similarity of the CS+ and US representations encodes the CS+–US association strength.


  1. Herry, C. et al. Switching on and off fear by distinct neuronal circuits. Nature 454, 600–606 (2008)

    Article  CAS  ADS  Google Scholar 

  2. Maren, S. & Quirk, G. J. Neuronal signalling of fear memory. Nat. Rev. Neurosci. 5, 844–852 (2004)

    Article  CAS  Google Scholar 

  3. Quirk, G. J., Armony, J. L. & LeDoux, J. E. Fear conditioning enhances different temporal components of tone-evoked spike trains in auditory cortex and lateral amygdala. Neuron 19, 613–624 (1997)

    Article  CAS  Google Scholar 

  4. Chowdhury, N., Quinn, J. J. & Fanselow, M. S. Dorsal hippocampus involvement in trace fear conditioning with long, but not short, trace intervals in mice. Behav. Neurosci. 119, 1396–1402 (2005)

    Article  Google Scholar 

  5. Quirk, G. J., Repa, C. & LeDoux, J. E. Fear conditioning enhances short-latency auditory responses of lateral amygdala neurons: parallel recordings in the freely behaving rat. Neuron 15, 1029–1039 (1995)

    Article  CAS  Google Scholar 

  6. Ghosh, K. K. et al. Miniaturized integration of a fluorescence microscope. Nat. Methods 8, 871–878 (2011)

    Article  CAS  Google Scholar 

  7. Ziv, Y. et al. Long-term dynamics of CA1 hippocampal place codes. Nat. Neurosci. 16, 264–266 (2013)

    Article  CAS  Google Scholar 

  8. Chen, T. W. et al. Ultrasensitive fluorescent proteins for imaging neuronal activity. Nature 499, 295–300 (2013)

    Article  CAS  ADS  Google Scholar 

  9. Garner, A. R. et al. Generation of a synthetic memory trace. Science 335, 1513–1516 (2012)

    Article  CAS  ADS  Google Scholar 

  10. Gore, F. et al. Neural representations of unconditioned stimuli in basolateral amygdala mediate innate and learned responses. Cell 162, 134–145 (2015)

    Article  CAS  Google Scholar 

  11. Namburi, P. et al. A circuit mechanism for differentiating positive and negative associations. Nature 520, 675–678 (2015)

    Article  CAS  ADS  Google Scholar 

  12. Senn, V. et al. Long-range connectivity defines behavioral specificity of amygdala neurons. Neuron 81, 428–437 (2014)

    Article  CAS  Google Scholar 

  13. Romanski, L. M., Clugnet, M. C., Bordi, F. & LeDoux, J. E. Somatosensory and auditory convergence in the lateral nucleus of the amygdala. Behav. Neurosci. 107, 444–450 (1993)

    Article  CAS  Google Scholar 

  14. Wolff, S. B. et al. Amygdala interneuron subtypes control fear learning through disinhibition. Nature 509, 453–458 (2014)

    Article  CAS  ADS  Google Scholar 

  15. Blair, H. T. et al. Unilateral storage of fear memories by the amygdala. J. Neurosci. 25, 4198–4205 (2005)

    Article  CAS  Google Scholar 

  16. Ciocchi, S. et al. Encoding of conditioned fear in central amygdala inhibitory circuits. Nature 468, 277–282 (2010)

    Article  CAS  ADS  Google Scholar 

  17. Erlich, J. C., Bush, D. E. & Ledoux, J. E. The role of the lateral amygdala in the retrieval and maintenance of fear-memories formed by repeated probabilistic reinforcement. Front. Behav. Neurosci. 6, 16 (2012)

    Article  Google Scholar 

  18. Kong, E., Monje, F. J., Hirsch, J. & Pollak, D. D. Learning not to fear: neural correlates of learned safety. Neuropsychopharmacology 39, 515–527 (2014)

    Article  Google Scholar 

  19. Mankin, E. A. et al. Neuronal code for extended time in the hippocampus. Proc. Natl Acad. Sci. USA 109, 19462–19467 (2012)

    Article  CAS  ADS  Google Scholar 

  20. Blair, H. T., Schafe, G. E., Bauer, E. P., Rodrigues, S. M. & LeDoux, J. E. Synaptic plasticity in the lateral amygdala: a cellular hypothesis of fear conditioning. Learn. Mem. 8, 229–242 (2001)

    Article  CAS  Google Scholar 

  21. Bishop, C. M. Pattern Recognition and Machine Learning Vol. 1 (Springer, 2007)

  22. Rodrigues, S. M., Schafe, G. E. & LeDoux, J. E. Molecular mechanisms underlying emotional learning and memory in the lateral amygdala. Neuron 44, 75–91 (2004)

    Article  CAS  Google Scholar 

  23. Sotres-Bayon, F. & Quirk, G. J. Prefrontal control of fear: more than just extinction. Curr. Opin. Neurobiol. 20, 231–235 (2010)

    Article  CAS  Google Scholar 

  24. Goosens, K. A., Hobin, J. A. & Maren, S. Auditory-evoked spike firing in the lateral amygdala and Pavlovian fear conditioning: mnemonic code or fear bias? Neuron 40, 1013–1022 (2003)

    Article  CAS  Google Scholar 

  25. Danielson, N. B. et al. Sublayer-specific coding dynamics during spatial navigation and learning in hippocampal area CA1. Neuron 91, 652–665 (2016)

    Article  CAS  Google Scholar 

  26. Rubin, A., Geva, N., Sheintuch, L. & Ziv, Y. Hippocampal ensemble dynamics timestamp events in long-term memory. eLife 4, e12247 (2015)

    Article  Google Scholar 

  27. Han, J. H. et al. Neuronal competition and selection during memory formation. Science 316, 457–460 (2007)

    Article  CAS  ADS  Google Scholar 

  28. Rashid, A. J. et al. Competition between engrams influences fear memory formation and recall. Science 353, 383–387 (2016)

    Article  CAS  ADS  Google Scholar 

  29. Rogerson, T. et al. Molecular and cellular mechanisms for trapping and activating emotional memories. PLoS One 11, e0161655 (2016)

    Article  Google Scholar 

  30. Paton, J. J., Belova, M. A., Morrison, S. E. & Salzman, C. D. The primate amygdala represents the positive and negative value of visual stimuli during learning. Nature 439, 865–870 (2006)

    Article  CAS  ADS  Google Scholar 

  31. Belova, M. A., Paton, J. J., Morrison, S. E. & Salzman, C. D. Expectation modulates neural responses to pleasant and aversive stimuli in primate amygdala. Neuron 55, 970–984 (2007)

    Article  CAS  Google Scholar 

  32. Rumpel, S., LeDoux, J., Zador, A. & Malinow, R. Postsynaptic receptor trafficking underlying a form of associative learning. Science 308, 83–88 (2005)

    Article  CAS  ADS  Google Scholar 

  33. Sah, P., Westbrook, R. F. & Lüthi, A. Fear conditioning and long-term potentiation in the amygdala: what really is the connection? Ann. NY Acad. Sci. 1129, 88–95 (2008)

    Article  CAS  ADS  Google Scholar 

  34. Fadok, J. P., Dickerson, T. M. & Palmiter, R. D. Dopamine is necessary for cue-dependent fear conditioning. J. Neurosci. 29, 11089–11097 (2009)

    Article  CAS  Google Scholar 

  35. Johansen, J. P. et al. Hebbian and neuromodulatory mechanisms interact to trigger associative memory formation. Proc. Natl Acad. Sci. USA 111, E5584–E5592 (2014)

    Article  CAS  Google Scholar 

  36. Gütig, R. & Sompolinsky, H. The tempotron: a neuron that learns spike timing-based decisions. Nat. Neurosci. 9, 420–428 (2006)

    Article  Google Scholar 

  37. Frémaux, N. & Gerstner, W. Neuromodulated spike-timing-dependent plasticity, and theory of three-factor learning rules. Front. Neural Circuits 9, 85 (2016)

    Article  Google Scholar 

  38. Pawlak, V., Wickens, J. R., Kirkwood, A. & Kerr, J. N. Timing is not everything: neuromodulation opens the STDP gate. Front. Synaptic Neurosci. 2, 146 (2010)

    Article  Google Scholar 

  39. Hinton, G. The ups and downs of Hebb synapses. PsycARTICLES 44, 10–13 (2003)

    Google Scholar 

  40. Johansen, J. P., Tarpley, J. W., LeDoux, J. E. & Blair, H. T. Neural substrates for expectation-modulated fear learning in the amygdala and periaqueductal gray. Nat. Neurosci. 13, 979–986 (2010)

    Article  CAS  Google Scholar 

  41. Courtin, J. et al. Prefrontal parvalbumin interneurons shape neuronal activity to drive fear expression. Nature 505, 92–96 (2014)

    Article  ADS  Google Scholar 

  42. Raybuck, J. D. & Lattal, K. M. Double dissociation of amygdala and hippocampal contributions to trace and delay fear conditioning. PLoS One 6, e15982 (2011)

    Article  CAS  ADS  Google Scholar 

  43. Thévenaz, P., Ruttimann, U. E. & Unser, M. A pyramid approach to subpixel registration based on intensity. IEEE Trans. Image Process. 7, 27–41 (1998)

    Article  ADS  Google Scholar 

  44. Mukamel, E. A., Nimmerjahn, A. & Schnitzer, M. J. Automated analysis of cellular signals from large-scale calcium imaging data. Neuron 63, 747–760 (2009)

    Article  CAS  Google Scholar 

  45. Betley, J. N. et al. Neurons for hunger and thirst transmit a negative-valence teaching signal. Nature 521, 180–185 (2015)

    Article  CAS  ADS  Google Scholar 

  46. Rigotti, M. et al. The importance of mixed selectivity in complex cognitive tasks. Nature 497, 585–590 (2013)

    Article  CAS  ADS  Google Scholar 

  47. George Paxinos, K. B. J. F. The Mouse Brain in Stereotaxic Coordinates Vol. 2 (2001)

  48. Singec, I., Knoth, R., Ditter, M., Volk, B. & Frotscher, M. Neurogranin is expressed by principal cells but not interneurons in the rodent and monkey neocortex and hippocampus. J. Comp. Neurol. 479, 30–42 (2004)

    Article  CAS  Google Scholar 

  49. Maren, S., Yap, S. A. & Goosens, K. A. The amygdala is essential for the development of neuronal plasticity in the medial geniculate nucleus during auditory fear conditioning in rats. J. Neurosci. 21, RC135 (2001)

    Article  CAS  Google Scholar 

Download references


G. Venkatraman, B. Ahanonu, J. Li, B. Rossi, C. Herry, S. Ciocchi and J. Bacelo provided technical assistance. We appreciate Swiss National Science Foundation (B.F.G.), Swiss National Science Foundation, Ambizione (J.G.), US National Science Foundation (L.J.K.), Stanford University (L.J.K., J.D.M.), Simons Foundation (L.J.K.), and Helen Hay Whitney Foundation (M.C.L.) fellowships. A.L. received support from the Swiss National Science Foundation, Novartis Research Foundation, and an ERC Advanced grant. M.J.S. received support from HHMI and DARPA.

Author information

Authors and Affiliations



B.F.G. designed experiments. B.F.G., J.G.P. and J.G. established the Ca2+ imaging protocol and performed experiments. B.F.G., P.E.J, A.L. and M.J.S. designed analyses. B.F.G. and J.D.M. analysed data. J.A.L., L.J.K., J.D.M. and M.C.L. provided software code and advised on analyses. J.Z.L. constructed virus. F.G. and A.L. provided electrophysiological data. B.F.G. and M.J.S. wrote the paper. J.G., A.L. and all authors edited the paper. A.L. and M.J.S. supervised the research.

Corresponding author

Correspondence to Mark J. Schnitzer.

Ethics declarations

Competing interests

M.J.S. is a scientific co-founder of Inscopix, Inc., which produces the miniature fluorescence microscope used in this study.

Additional information

Reviewer Information Nature thanks V. Bolshakov and the other anonymous reviewer(s) for their contribution to the peer review of this work.

Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data figures and tables

Extended Data Figure 1 Mouse preparation for Ca2+ imaging in excitatory BLA neurons.

a, Coronal slice of a mouse brain showing expression in the BLA of the GCaMP6m Ca2+ indicator, 5 weeks after injection of the AAV2/5-CaMK2a-GCaMP6m virus. Scale bar, 1 mm. b, Schematic of a coronal mouse brain section shown with the reconstructed positions (dashed red lines) of the microendoscope implants in the BLA, for the 12 mice subject to the experimental protocol of Fig. 1c. The focal planes for in vivo Ca2+ imaging were 77–181 μm below the indicated implant positions, as determined by computational modelling of the microendoscope optical pathway, using the empirically determined value of the back focal length. The optical focal plane often spanned ventral parts of lateral amygdala and dorsal parts of the basal amygdala, hence the use of the term basal and lateral amygdala (BLA) throughout. AP, anterior posterior; Ctx, piriform cortex. Scale bar, 1 mm. The mouse brain section has been reproduced with permission from ref. 47. c, Top, wide-field fluorescence image of BLA tissue acquired through an implanted microendoscope, 6 weeks after injection of the AAV2/5-CaMK2a-GCaMP6m virus. The outer fibre tract enclosing the BLA does not express GCaMP6m and appears as a vertical dark stripe in the centre of the field-of-view. The dashed box shows the position of the camera’s field-of-view, which we positioned over the BLA by using the fibre tract as a reference marker. Bottom, the same image but with the boundaries of the BLA and endopiriform nucleus (Epn) marked in green and black dashed lines, respectively. Scale bar, 0.2 mm. df, Coronal section of a mouse brain showing, inhibitory neurons in the BLA immuno-labelled with a monoclonal anti-GAD67 antibody (d), neurons expressing GCaMP6m under the control of the Camk2a promoter (e) and the overlay of the images in d and e (f). Red arrows in d and e mark GAD67+ interneurons that are not expressing GCaMP6m. Scale bars, 20 μm. gi, Coronal brain section showing excitatory neurons in the BLA immunolabelled using a polyclonal anti-neurogranin (NRGN) antibody48 (g), neurons expressing GCaMP6m (h) and an overlay of the images in g and h (i), showing that the set of NRGN+ excitatory neurons (labelled red) strongly overlap with the set of cells expressing GCaMP6m (labelled green). Scale bars, 20 μm.

Extended Data Figure 2 Stimuli of neutral, positive, and negative valence activate sparse, largely distinct, spatially intermingled subsets of neurons in the BLA.

a, A miniature fluorescence microscope enabled large-scale neural Ca2+ imaging in the BLA of awake behaving mice as we presented stimuli of different valences to the animals. b, Candidate cells identified using an automated cell sorting routine7,44 were easily segregated into those with sizes, morphologies and Ca2+ activity traces (grey traces, individual activity transients; black traces, mean waveforms) characteristic of individual neurons (left), and those that were obviously not neurons (right). For the 4–10% of candidates with less common characteristics, we accepted only those that were plainly cells by human visual scrutiny. We verified every cell included in the analyses by visual inspection. c, An example cell map in the BLA, as determined from a Ca2+ imaging dataset 28 min in duration. Colours indicate subsets of BLA neurons that responded significantly to rewards (light blue), electric foot shocks (green), eyelid shocks (yellow), or neutral tones (red). Scale bar, 20 μm. P ≤ 0.01, rank-sum test, comparing evoked Ca2+ signals to baseline levels. d, Ca2+ responses of six example neurons in the same mouse after the delivery of individual water rewards (left), eyelid shocks (middle) or foot-shocks (right). Grey traces show Ca2+ responses from eight individual trials. Black traces show the mean responses. e, Percentages of cells (n = 1,251 neurons in total from 8 mice) with significant Ca2+ responses to the four different stimuli (threshold for a significant response: P ≤ 0.01, comparing evoked versus baseline Ca2+ levels for n = 8 presentations of the stimulus; Wilcoxon rank-sum test). Error bars denote s.e.m. f, Cumulative probability distributions, each determined as a mean over 8 mice (1,251 total cells), of the centroid separations between all pairs of cells in each mouse (dashed black curve), and between pairs of cells that both had significant responses to one of the four different stimuli (coloured curves). Inset, the corresponding probability densities. g, Percentages of all neurons (n = 8 mice; 1,251 cells in total) that had significant responses to each of the two stimuli in each pair listed on the vertical axis. Dashed orange line indicates expected levels of overlap due to random chance. *P < 0.05, **P < 0.01, comparing actual percentages versus those determined from datasets in which we randomly shuffled the identities of the cells (1,000 random shuffles; Wilcoxon signed-rank test). h, Mahalanobis PVDs between the ensemble neural representations of the two stimuli of each pair listed on the vertical axis. All PVDs are normalized to the PVD between the neural representations of eyelid-shock and foot-shock. Pairs of stimuli with smaller PVDs have ensemble neural representations of greater similarity than pairs with larger PVDs. Dashed orange line indicates PVDs between ensembles in which we randomly shuffled the identities of the cells (1,000 random shuffles). *P < 0.05, **P < 0.01, comparing actual PVD values to those determined for the shuffled datasets, Wilcoxon signed-rank test. Data are based on the same 1,251 cells as in eg. Error bars in g and h denote s.e.m. i, 18 sets of fluorescence Ca2+ traces, showing evoked responses to presentations of the CS+ and CS from 18 example neurons before fear conditioning. Light grey traces show the individual responses of the cells to five stimulus presentations; black traces are average responses. Traces were downsampled to 5 Hz to aid visualization.

Extended Data Figure 3 Unilateral implantation of a microendoscope does not alter conditioned freezing; bilateral implantation minimally alters conditioned freezing without affecting locomotion.

a, Traces of locomotor activity across an entire (22 min) habituation session (day 1), for one example mouse in each of the three experimental groups indicated. Scale bar, 5 cm. b, Total distance travelled (left), locomotor speed (middle) and acceleration (right) for the three groups of mice during the day 1 habituation session. No significant differences between the three experimental groups (no-implant control (12 mice); unilateral implant (12 mice); bilateral implant (10 mice)) (one-way Kruskal–Wallis test; degrees of freedom: dfgroup = 2, dferr = 31, dftotal = 33 for 3 groups and 34 total mice; χ2 = 10–12; P ≥ 0.05 for all three locomotor parameters). c, d, Time mice spent freezing before conditioning (days 1, 2) in response to 5 presentations of the CS (c) and the CS+ (d), in no implant controls (12 mice), unilateral implant (12 mice), bilateral implant (10 mice), and bilateral implant plus muscimol BLA injection before the day 3 conditioning session (8 mice). No significant differences in freezing time (one-way Kruskal–Wallis test; dfgroup = 3, dferr = 42, dftotal = 45 for 4 groups and 42 total mice; χ2 = 10.2 and 11.8 for CS+ and CS, P ≥ 0.05 for both CS+ and CS). e, f, Time mice spent freezing after conditioning (days 4–6) in response to 4 presentations of the CS (e) and during 3 sets each comprising 4 presentations of the CS+ (f) in the same 42 mice as in c and d. *P = 0.005 (Wilcoxon signed-rank test; bilateral muscimol group versus control; significance threshold = 0.02 after Dunn–Šidák correction for 3 comparisons). Data are consistent with a study showing the necessity of BLA for auditory fear conditioning49 and further demonstrate that the BLA we are imaging are functional and necessary for the behaviour. g, Time mice (n = 12) spent freezing during the 20–180 s inter-stimulus intervals (ISI) after either a CS+ or CS presentation. CS+ and CS freezing values are averages over the numbers of stimulus presentations shown in Fig. 1c. After fear conditioning, CS-evoked freezing levels were above those during the inter-stimulus intervals, indicating the CS did not serve as a learned safety signal. Data in b–g are mean ± s.e.m.

Extended Data Figure 4 Ca2+ transient responses of individual BLA neurons to CS presentations closely resemble expectations based on electrical recordings of these responses.

To assess whether fluorescence Ca2+ imaging in the BLA captured similar forms of neural activity as previous extracellular electrical recordings in this brain area, we compared the responses of individual neurons to CS presentations, as observed using the two recording modalities in two different sets of mice presented with the same CS stimuli. Across the two datasets, there was close agreement between the shapes of the empirically determined Ca2+ transient waveforms and the expected waveforms based on the electrically recorded CS-evoked spiking responses. a, We took a recording of CS-evoked spiking activity from an individual BLA cell (left), convolved the spike train with a decaying exponential function (700 ms time constant) to account for the kinetics of the GCaMP6m indicator (middle), and subtracted the baseline fluorescence level to yield a predicted CS-evoked Ca2+ fluorescence signal (ΔF/F) whose waveform shape closely matched the actual CS-evoked Ca2+ fluorescence signal of a BLA cell that we had monitored using the miniature microscope (right). b, Six additional examples of the CS-evoked spiking responses in individual BLA neurons, as monitored via extracellular electrical recordings (black traces). From these spike trains, the same approach as shown in a was used to predict the Ca2+ fluorescence signals that these cells would produce (red traces), and these predictions were compared to the actual CS-evoked Ca2+ fluorescence signals of another six BLA cells studied by Ca2+ imaging with similar responses (blue traces).

Extended Data Figure 5 Precise spatial registration of the Ca2+ imaging datasets from different behavioural sessions allows unambiguous tracking of individual cells across multiple days.

a, Using the spatial filters provided for each neuron by the automated cell sorting algorithm7,44, maps of all active cells detected in the BLA on each day of the study were made. Standard methods of image alignment43 were used to register these maps across the different days. Approximately 50% of all neurons observed across the entire experiment were detected as active on individual days. a, Example maps of active BLA cells from three mice on the first (left), third (middle), and last (right) day of the 6-day experimental protocol (Fig. 1c). Circles indicate cells that were active in only one of the three recordings (grey), on two of the three days (blue), or on all three days (red). Scale bars, 30 μm. The maximum number of active cells seen in one session was 192. b, Thresholded spatial filters from three example cells registered across the 6-day experimental protocol. Green asterisks indicate the centroid position of each cell on day 1. Blue asterisks mark the centroid positions on subsequent days. Scale bar, 10 μm. c, Five examples of neighbouring cells detected via their activity patterns on different days. Two individual cells are clearly discernible in each case. Scale bar, 10 μm. d, Cumulative histogram of distances between the centroids of all pairs of cells detected within the same imaging session, plotted with a logarithmic scale. Inset, magnified view of the dashed box area. No pairs of cells were separated by <6 μm. e, Cumulative histogram of distances between the centroids of all pairs of active cells registered as being the same neuron seen in different sessions. Inset, magnified view for y-axis values >97%. Because the worst-case alignment error of the image registration algorithm was 1.5 μm, as determined by bootstrap analysis7, and as all pairs of anatomically distinct cells were separated by ≥6 μm (d), cell pairs separated by <4.5 μm were nearly guaranteed to be the same neuron seen on two different occasions. This yielded the worst-case estimate that >99.7% of all cell pairs registered as being the same cells were correctly assigned the same identity. This estimate is conservative in that the image registration errors were usually <1 μm. f, Probability that an active neuron detected in one imaging session will also be active in a subsequent session, for all 3,655 neurons (black) and for CS-responsive neurons (grey). Inset, probability that a cell detected on any day in the study was present in each of the imaging sessions, for all 3,655 neurons (black), the CS+-responsive neurons (red), and the CS-responsive neurons (blue). These probabilities were constant throughout and statistically indistinguishable between the three groups of cells examined for all days and all mice (49 ± 2% of all cells were active each day; two-way repeated measures ANOVA; degrees of freedom: dfmice = 11, dfgroup = 2, dfinteraction = 22, dferr = 180, dftotal = 215 for 6 days, 3 groups of cells and 12 mice; χ2 = 0.7–1.4; P > 0.05 for all). g, Number of neurons detected in each mouse was stable across all days (152 ± 14 cells per day; n = 12 mice; one-way Friedman test; dfdays = 5, dferr = 55, dftotal = 71 for 6 days and 12 mice; χ2 = 5.9; P = 0.31). h, Percentages of all 3,655 cells in the study that were detected in 1–6 sessions. Data in f–h are mean ± s.e.m.

Extended Data Figure 6 Conditioning induces bi-directional changes in CS-evoked responses.

a, b, Contrary to the predictions of the cellular, Hebbian model of fear learning, conditioning induced substantial bi-directional changes in the CS+-evoked responses of cells that responded and those that did not respond to the US. Notably, a preponderance of cells that responded to both the CS+ and US before training had decreased CS+-evoked responses after training (a). Furthermore, many cells with potentiated CS+-evoked responses after training were not US-responsive (b). Ensemble level analyses showed that cells with up- and downregulated responses made equally important contributions to the learning-induced changes in ensemble neural coding (Fig. 3e). a, Percentages of CS-responsive cells that were also US-responsive (blue) and of CS+-responsive cells that were also US-responsive. The latter are further divided into cells that increased their CS+-evoked responses after training (maroon), those that underwent no significant changes in their CS+-evoked responses (pink), and those that decreased their CS+-evoked responses after training (red). b, Percentages of CS-responsive cells that were not US-responsive (blue) and of CS+-responsive cells that were not US-responsive. The latter are further divided into cells that increased their CS+-evoked responses after training (maroon), those that underwent no significant changes in their CS+-evoked responses (pink), and those that decreased their CS+-evoked responses after training (red). All data are from the same 12 mice and denote mean ± s.e.m.

Extended Data Figure 7 BLA ensembles provide sufficient information to decode the CS, and the decoding accuracy improves with successive tone presentations in a series of tones.

a, Left, a three-way decoder has three possible outputs (CS+, CS and baseline) and hence different categories of possible errors. When a decoder makes a type A error, it outputs the wrong CS (for example, CS+ instead of CS). When a decoder makes a type B error, it fails to distinguish a CS presentation from baseline activity. Right, when the activity traces of all neurons were used to train the decoders, the decoders yielded the correct answer on 97 ± 1% (mean ± s.e.m.) of all trials from a testing set comprising equal numbers of samples of each type. The success rate was 90 ± 2% when the decoders were trained using only those cells with statistically significant responses to at least one of the CS types. b, For trials that were incorrectly decoded, the pie charts show the proportions of the two types of error, for decoders trained on the activity traces of all neurons (left), and using only neurons with statistically significant responses to at least one of the two CS types (right). c, Type A errors declined sharply during the first 5 of the 25 CS tone pulses (black), both before (left), and after (right) conditioning. After conditioning, as the 25 tone pulses proceeded, the mice increasingly distinguished between the CS (blue) and the CS+ (red), as seen by the differences in evoked freezing behaviour. Data are mean ± s.e.m. d, Schematic showing how we extracted the principal components (PCs) of the BLA ensemble responses to CS tone presentations. Dashed box encloses two PCs, used in e for illustration. e, Plots of the first two PCs, determined as in d for four example mice, illustrate that the ensemble responses to the CS+ (red) and the CS (blue) were generally distinguishable. Black stars mark the first out of 25 tone pulses for each CS presentation and illustrate that the initial tones in the series were generally the hardest to categorize correctly.

Extended Data Figure 8 Fisher linear discriminant analysis of BLA population activity.

a, Notional schematic showing the Mahalanobis distance and the discrimination boundary of a Fisher linear discriminant analysis (FLDA) decoder (black dotted line), which discriminates the multi-dimensional, neural ensemble responses to CS presentations from the activity patterns during baseline conditions. For simplicity, the schematic shows a hypothetical case in which the ensemble consisted of only two neurons, but the basic principles readily apply to larger ensembles. For a given set of training data, the Fisher decoder provides the a posteriori probability that a representative data sample will be correctly categorized. b, Example histogram from one mouse showing BLA ensemble responses (day 1) to CS presentations, normalized and projected onto the dimension of maximal discriminability. Dashed vertical line marks the classification boundary of the Fisher linear decoder, dividing those ensemble responses classified as baseline from those representing a CS. The separation between the two peaks in the histogram is an empirical estimate of the Mahalanobis distance, which is a multi-dimensional generalization of the discriminability index, d’, used in statistics21. c, Mean decoding performance as a function of the number of cells used for training the decoder (open circles) and corresponding parametric fits to a sigmoid function. When the training and testing data came from the same day (black curve), performance asymptotically approached near perfect decoding when more than ~100 cells were used. When the training and testing data came from different days (red curve), our datasets were not large enough to approach the asymptote. However, the sigmoidal fit suggests that the asymptotic performance of time-lapse decoders would be around 90% in cases with more than 120 cells. Shading indicates s.e.m. d, Decoding performance of time-lapse CS decoders, as a function of the elapsed time between the day on which the training dataset was acquired and the day on which the testing dataset was acquired (n = 12 mice). Despite declining re-occurrence probabilities of the cells as a function of elapsed time (Extended Data Fig. 5f), decoding performance remained stable for time-lapse intervals of 1–5 days. Data are mean ± s.e.m.

Extended Data Figure 9 The Mahalanobis distance quantifies the discriminability of two sets of ensemble responses and takes into account the mean and covariance of each response set.

a, Left, schematic of two sets of ensemble neural responses (blue and red clouds of data points), illustrated for a hypothetical case in which the ensemble consists of only two neurons. The Euclidean distance (grey line) between the means of the two distributions does not take into account the degree to which the ensemble neural responses are variable from trial to trial. Right, to characterize the differentiability of the two response sets in a way that takes into account neural variability, the Mahalanobis distance (M) between the two distributions was determined. To do this, the covariance matrix of the ensemble neural responses (Σ) was used to map the data points into a space in which the distributions have unity variance in all directions. M is equivalent to the distance between the means of the two resulting distributions. b, A change in the Mahalanobis PVD can be due to changes in the means of the two sets of ensemble responses, changes in response variability, or both. Schematic illustrates these two different ways in which the sets of ensemble responses can become more or less differentiable. Top row shows a pair of cases in which changes in the mean ensemble responses dominate the change in the PVD. Bottom row shows a pair of cases in which changes in response variability dominate the change in the PVD. c, The total change in the CS+–US PVD (red curve) induced by learning was divided into two components: a component due to changes in the mean CS+-evoked response (cyan curve) and a component due to changes in the variability of the CS+-evoked responses (black curve). After conditioning, the CS+-evoked responses became less variable (black curve) but also more similar to the US-evoked ensemble responses (cyan curve). The latter effect substantially outweighed the former, leading to a net ~32% decline (red curve) in the differentiability of the CS+- and US-evoked responses, as quantified by the net decrease in the Mahalanobis distance. Thin lines show the values from each of 12 individual mice. Thick lines show the mean values. Error bars are s.e.m.

Extended Data Figure 10 Procedure for computational rescue of the CS+ decoders.

Unlike time-lapse CS decoders, which worked well across all 6 days of the experiment, time-lapse CS+ decoders did not work well when the training and testing datasets came from a pair of days that spanned across the conditioning session (Fig. 3b). This failure mode for the CS+ decoders arises from the learning-induced changes in the ensemble representation of the CS+ (Fig. 3c, d). However, by extrapolating the changes in the CS+ representation that occur during the training session on day 3, we could predict the much greater, subsequent changes in the CS+ representation that occur before the next session on day 4 and thereby rescue the failures of the time-lapse CS+ decoders. This figure schematizes the procedure for the computational rescue. a, Schematic illustration of how conditioning-induced changes in CS+-evoked ensemble activity (light and dark red dots) can impair the performance of a time-lapse decoder trained on data from before fear conditioning and tested on data from after conditioning. b, Using five main steps, we computationally simulated the changes in the CS+-representation that occurred during post-training consolidation, by extrapolating by a factor, q, the much smaller changes in the CS+-representation that occurred during the day 3 training session. c, To determine the optimal value of the extrapolation factor q, we simulated the post-training changes in the CS+-representation by computationally adjusting the CS+ population vectors in increments of ΔA, the modest change in coding that occurred on day 3. Increments of 3–5 times ΔA were optimal, in that they best rescued the capabilities of two-way decoders trained on either of days 1 or 2 to detect a CS+ presentation when tested on data from after training (days 4–6), or vice versa. (Each datum shows the mean ± s.e.m. decoding performance, averaged across 12 mice and the 12 possible pair-wise combinations per mouse of one pre- and one post-training day.) The same analysis of the CS representation scarcely yielded any change in decoding performance, because the effects of training (ΔA) for the CS were negligible. Decoders trained on temporally shuffled data (1,000 shuffles; grey curve) and those based only on cells with up- (green) or downregulated (purple) responses to the CS+ after training performed less successfully than decoders based on all cells (brown).

Supplementary information

Supplementary Information

This file contains Supplementary Tables 1-2, a Supplementary Note and Supplementary References. (PDF 376 kb)

PowerPoint slides

Rights and permissions

Reprints and Permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Grewe, B., Gründemann, J., Kitch, L. et al. Neural ensemble dynamics underlying a long-term associative memory. Nature 543, 670–675 (2017).

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI:

This article is cited by


By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.


Quick links

Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing