Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2023 Jun 21;111(12):1952-1965.e5.
doi: 10.1016/j.neuron.2023.03.011. Epub 2023 Apr 3.

Emergence of a predictive model in the hippocampus

Adam M P Miller  1 , Alex D Jacob  2 , Adam I Ramsaran  2 , Mitchell L De Snoo  3 , Sheena A Josselyn  4 , Paul W Frankland  5
Affiliations

Emergence of a predictive model in the hippocampus

Adam M P Miller et al. Neuron. .

Abstract

The brain organizes experiences into memories that guide future behavior. Hippocampal CA1 population activity is hypothesized to reflect predictive models that contain information about future events, but little is known about how they develop. We trained mice on a series of problems with or without a common statistical structure to observe how memories are formed and updated. Mice that learned structured problems integrated their experiences into a predictive model that contained the solutions to upcoming novel problems. Retrieving the model during learning improved discrimination accuracy and facilitated learning. Using calcium imaging to track CA1 activity during learning, we found that hippocampal ensemble activity became more stable as mice formed a predictive model. The hippocampal ensemble was reactivated during training and incorporated new activity patterns from each training problem. These results show how hippocampal activity supports building predictive models by organizing new information with respect to existing memories.

Keywords: context; hippocampus; inference; integration; latent state; learning; memory; prediction; retrieval; schema.

PubMed Disclaimer

Conflict of interest statement

Declaration of interests The authors declare no competing interests.

Figures

Figure 1.
Figure 1.
Efficient learning of structured versus unstructured problems. (A) Mice were trained and tested in an operant chamber with a nose-poke port, reward-delivery hopper, and a speaker that delivered different tones. Each trial consisted of a nose-poke followed by head entry into the hopper. Upon head entry, a tone played followed by a potential reward delivery. (B) Two groups of mice were trained on a series of tone discrimination problems. In the Structured training group, all of the rich tones fell within a continuous band of frequencies. (C) In the unstructured group, the rich tones were more evenly distributed. The first and last problem were the same for both groups (black arrows). (D) Preference for the rich tone is shown in terms of the standardized mean difference between how long mice waited in response to the rich tone and the poor tone. Preferences are shown for the first and last day of training on every problem. Mice were trained to the same criterion (longer wait times on rich tone trials than poor tone trials as determined by a significant T-test, alpha level = 0.01) on every problem and then were given one additional day of training (i.e., the last day). The Structured training group showed better discrimination on the first day of problems later in learning compared to the first problem. (E) The Unstructured training group did not improve on its first-day performance and typically showed significantly worse performance. (F) The number of days required to reach criterion. The Structured training group learned new problems faster later in learning. (G) The learning rate of the Unstructured training group failed to improve. Direct comparisons between the two groups on the first and last problem revealed that the Structured training group (H) showed higher initial discrimination on the last training problem, (I) learned the last problem in fewer days, and (J) learned less new information on the last problem compared to the Unstructured training group. Error bars show standard error of the mean. *p<0.05, **P<0.01, ***P<0.001
Figure 2.
Figure 2.
Integrating structured experiences into a generalized model. (A) The mean wait time responses to every tone are shown for all subjects in every probe test of the Structured training group. Light lines show the responses of individual subjects. Probe 0 is the pre-probe administered before any tone training. Subsequent probes were administered the day after completing the corresponding problem. The rich and poor tone frequencies of the prior problem are shown as red and grey lines, respectively. (B) Same as A, but for the Unstructured training group showing no integration of new information into previous memories. (C) Correlation matrices showing the mean of all subject correlations between every pair of probe tests for the Structured training group (top) and Unstructured training group (bottom). (D) The correlation between every pair of sequential probe tests is shown for the Structured training group. (E) Same as D, but for the Unstructured training group. (F) The correlation between the probes before and after the two common problems (problem 1 and problem 6) are shown for both groups. Error bars show standard error of the mean. *p<0.05, **P<0.01, ***P<0.001
Figure 3.
Figure 3.
Predictive models guide new learning. (A) Memory for the reward values of the rich and poor tones used in the prior problem can be measured during the probe test one day after training. All wait times from two example sessions (the last day of training on the first problem and the probe test one day later) are shown from one mouse. Grey circles indicate the wait times on individual trials. We quantified the memory for the prior problem in terms of the difference between the how long the mouse waited in response to the same tones during the subsequent probe test. Wait times were determined by fitting a curve to the observed waits from every trial. (B) The Structured training group and (C) the Unstructured training group showed significant memory for the prior problem during every probe test. Light lines show data from individual subjects. (D) The two groups also showed similar memory for each of the two common training problems. (E) Predictions for the reward values of future rich and poor tones can be measured during the probe test one day before the start of training. We measured predictions in the same way that we measured memory, but with the rich and poor tones from the upcoming (and not the prior) problem. Example data are from one subject during the 5th probe and the first day of training on the last problem. (F) The Structured training group accurately predicted the solutions to some problems later in learning after initially making inaccurate predictions. (G) The Unstructured training group showed inaccurate predictions throughout training. (H) The Structured training group developed superior predictions with training compared to the Unstructured training group. (I) The more accurate the prediction (measured during the preceding probe test), the stronger the preference for the rich tone on the first day of training, defined as the standardized mean difference between the wait times in response to the rich and poor tones. Large circles show condition means. (J) The more accurate the prediction, the stronger the preference for the rich tone on the last day of training. (K) The more accurate the prediction, the faster the subject learned the training problem. (L) The more accurate the prediction, the less new information was learned during a training problem, defined as a smaller change in the preference for the rich tone from the first to the last day of training. (M) More accurate predictions were associated with smaller learning-related changes to the memory, defined in terms of the correlation between the behavioral responses on probe tests before and after training (higher correlations indicate smaller changes). Error bars show standard error of the mean. *p<0.05, **P<0.01, ***P<0.001
Figure 4.
Figure 4.
Mice reactivate CA1 populations during structured, but not unstructured, training. (A) Cartoon of the custom miniature microscope implanted into the hippocampus. (B) Histological section used to identify the location of a GRIN lens implanted on the surface of CA1. (C) We tracked the activity of every neuron throughout training for both the Structured training and Unstructured training groups (see also Figure S3). The columns show the activity of one neuron over all 19 training sessions (7 probes and the first and last day of each of 6 problems). Neurons are colored according to the session in which they were first active. Rows show the composition of every session in terms of which cells were active. The composition is summarized in the outset bar chart. (D) An explanation and example of the population overlap computation, defined as the number of cells active in both sessions divided by the number active in either. (E) The mean observed overlap between every pair of imaging sessions for the Structured training and Unstructured training groups. Matrices showing excess overlap are shown in Figures S4D–E.
Figure 5.
Figure 5.
Mice reactivate CA1 activity patterns during Structured training. (A) We examined activity during the window between nose-poke and then end of the minimum wait (MW) period, occurring 2s after tone-on. (B) Three examples of neurons that were reliably active during discrete periods on every trial. (C) The mean activity is shown for every neuron imaged from each group, arranged by when during the trial it was most active. (D) The activity patterns of neurons that were active across any pair of probe sessions surrounding common problems (probes 0 & 1, 5 & 6) are shown. We examined the similarity of population activity across sessions by measuring the correlation between the activity patterns of individual cells in both sessions. The top row shows all neurons imaged from mice in the Structured training group that were active in both probe 0 and 1 (left side), and then neurons that were active in both probe 5 and 6 (right side). The bottom row shows neurons imaged from mice in the Unstructured training group. For each pair of probes, the neurons are sorted according to their activity in the left-side probe (probe 0 or probe 5), such that corresponding rows show the activity of the same neuron. Note the highly similar activity patterns between probes 5 and 6 in the Structured training group only. (E) The mean observed activity pattern similarity between every pair of imaging sessions for the Structured training and Unstructured training groups. Matrices showing excess activity similarity are shown in Figure S4F–G.
Figure 6.
Figure 6.
Learning-related neuronal activity is incorporated into a reactivated ensemble. (A) We computed a reactivation score that combined ensemble overlap and activity similarity measures. The reactivation probe session population during subsequent training problems is shown for the Structured training group (left) and the Unstructured training group (right). Light lines show data from individual subjects. (B) The reactivation of probe ensembles during subsequent common problems is shown for both groups. (C) The reactivation of probe ensembles during subsequent problems was predicted by first day discrimination accuracy, a behavioral measure of predictive memory retrieval. Large circles show condition means. (D) The reactivation of training session ensembles during the final probe session is shown for the Structured training group (left) and the Unstructured training group (right). (E) The reactivation of the common training problem ensembles during the final probe session is shown for both groups. (F) The reactivation of training ensembles during the final probe was predicted by the reactivation of probe ensembles during training. (G) Like D, but with the reactivation of probe ensembles during the final probe session. (H) Like E, but with probe ensembles. (I) The reactivation of probe ensembles during the final probe was predicted by the reactivation of probe ensembles during training problems. Error bars show standard error of the mean. *p<0.05, **P<0.01, ***P<0.001
See this image and copyright information in PMC

Comment in

References

    1. Kumaran D, Hassabis D, and McClelland JL (2016). What Learning Systems do Intelligent Agents Need? Complementary Learning Systems Theory Updated. Trends Cogn Sci 20, 512–534. 10.1016/j.tics.2016.05.004. - DOI - PubMed
    1. Tervo DGR, Tenenbaum JB, and Gershman SJ (2016). Toward the neural implementation of structure learning. Curr Opin Neurobiol 37, 99–105. 10.1016/j.conb.2016.01.014. - DOI - PubMed
    1. Behrens TEJ, Muller TH, Whittington JCR, Mark S, Baram AB, Stachenfeld KL, and Kurth-Nelson Z (2018). What Is a Cognitive Map? Organizing Knowledge for Flexible Behavior. Neuron 100, 490–509. 10.1016/j.neuron.2018.10.002. - DOI - PubMed
    1. Momennejad I. (2020). Learning Structures: Predictive Representations, Replay, and Generalization. Curr Opin Behav Sci 32, 155–166. 10.1016/j.cobeha.2020.02.017. - DOI - PMC - PubMed
    1. Morton NW, and Preston AR (2021). Concept formation as a computational cognitive process. Curr Opin Behav Sci 38, 83–89. 10.1016/j.cobeha.2020.12.005. - DOI - PMC - PubMed

Publication types

LinkOut - more resources