Why Do Memories Persist When the Brain Keeps Changing?

When we recall something that happened yesterday, our brain is no longer in the same state it was then. We have had further experiences, slept and become active again. Yet our memories retain a degree of consistency. What, within the brain, has to be preserved for a memory to persist?

Research using artificial hibernation by OIST and its collaborators offers a starting point for exploring this question. Even after changes involving the loss of many synapses in the mouse hippocampus, memory-related behavior was preserved following recovery. The researchers also observed that some clusters of memory-related synapses were more likely to remain. It is interesting to read artificial hibernation as an experiment that substantially changes a circuit and asks what survives. OIST research announcement ↗

Sleep research offers findings that connect with the same question. Norimoto, Ikegaya and colleagues showed that interfering with synaptic weakening associated with hippocampal ripples impaired subsequent new learning. The weakening was selective. This raises the possibility that adjusting activity that is no longer needed, or excessive connections, creates room for further learning. Norimoto et al., Science, 2018 ↗

Research in the mouse motor cortex has also shown that, during REM sleep, some newly formed spines are eliminated while others are strengthened and maintained. Circuits are not simply weakened uniformly during sleep. Might the selection of what to retain and what to update help reconcile memory maintenance with subsequent learning? Li et al., Nature Neuroscience, 2017 ↗

Connections are not the only things that change. Even in mice that stably perform a learned task, the relationship between individual neurons in the posterior parietal cortex and task information changes over time. Yet task information can still be read out from the population's activity. This shows that maintaining an ability cannot be fully explained by each individual cell continuing to respond in the same way. Driscoll et al., Cell, 2017 ↗

Analyzing these data, Rule and colleagues investigated the conditions under which information could be extracted from changing representations. They presented a model in which small, ongoing adjustments to the readout could compensate for the effects of those changes on performance. If the population carrying a representation and the circuit using it are updated together, the same judgments or behaviors might continue even as their underlying configuration changes. Rule et al., eLife, 2020 ↗

At the same time, we cannot ignore which connections remain. Hayashi-Takagi and colleagues showed that selectively shrinking spines potentiated during learning impaired an acquired motor skill. Manipulating spines associated with a different task did not produce the same impairment. In addition to asking how much a circuit has changed overall, we need to ask which of the relationships supporting learning have changed. Hayashi-Takagi et al., Nature, 2015 ↗

This is where I would like to bring dropout into the comparison. Temporarily making units unavailable during training can reduce reliance on particular combinations and improve generalization. Useful representations can be learned while experiencing conditions in which some components are unavailable. Srivastava et al., JMLR, 2014 ↗

The comparison raises a question: what kinds of robustness can be developed through exposure to variability? In the brain, too, we can consider mechanisms that reproduce learned distinctions and associations despite some degree of change, rather than simply preserving particular connections or activity states unchanged. Relatively stable connections that provide a structural foundation, information preserved at the population level and a readout that adapts to change are not mutually exclusive explanations.

What we want to follow is the difference between changes that allow memories to persist and changes that cause them to break down. If the same number of connections are weakened, will the outcome differ depending on whether they are selected randomly or targeted within a cluster associated with a particular learning experience? Is memory retained even without relearning, or is it re-established during recovery? These comparisons could help us investigate what supports consistency in memory at the levels of wiring, population representations and readout.

Artificial hibernation research is one way to approach this general question by introducing a substantial change. Considering it alongside everyday sleep and learning may allow us to revisit memory maintenance and the brain's capacity to keep learning as aspects of the same problem.