Our recollection of events is usually not like a replay of digital video from a security camera—a passive observation that faithfully reconstructs the spatial and sensory details of everything that happened. More often memory segments what we experience into a string of discrete, connected events. For instance, you might remember that you went for a walk before lunch at a given time last week without recalling the soda bottle strewn on the sidewalk, the crow cawing in the oak tree in your yard or the chicken salad sandwich you ate upon your return. Your mind designates a mental basket for “walk” and a subsequent bin for “lunch” that, once accessed, make many of these finer details available. This arrangement raises the question of how the brain performs such categorization.
A new study by neuroscientist Susumu Tonegawa of the Massachusetts Institute of Technology and his colleagues claims to have discovered the neural processing that makes this organization of memory into discrete units possible. The work has implications for understanding how humans generalize knowledge, and it could aid efforts to develop AI systems that learn faster.
A brain region called the hippocampus is critical for memory formation and also seems to be involved in navigation. Neurons in the hippocampus called “place” cells selectively respond to being in specific locations, forming a cognitive map of the environment. Such spatial information is clearly important for “episodic” (autobiographical rather than factual) memory. But so, too, are other aspects of experience, such as changing sensory input. There is evidence that neurons in the hippocampus encode sensory changes by altering the frequency at which they fire, a phenomenon termed “rate remapping.” According to research by neuroscientist Loren Frank of the University of California, San Francisco, and his colleagues, such changes may also encode information about where an animal has been and where it is going, enabling rate remapping to represent trajectories of travel.
Besides coding continuously changing variables, whether sensory inputs or route trajectories, some imaging studies previously suggested that the brain also processes experience as segmented events. But exactly how it achieves this process at a neural level was not known. In the new study, published last week in Nature Neuroscience, the team—led by Chen Sun, a graduate student in Tonegawa’s lab—devised a task that attempted to disentangle the discrete, segmented nature of events from the continuously changing spatial and sensory details of moment-to-moment experience. The researchers trained mice to run around a square track. After doing four laps, the animals were rewarded with a sweet treat. They visited the reward box after every lap, segmenting each trial into four “events” (with the reward defining the end of a trial). Each lap traversed the same route, so sensory and location information was constant from one event to the next, allowing the researchers to attribute brain activity differences to what did change: the laps, or events.
The researchers recorded activity in hundreds of hippocampal cells while the mice performed this task and found that around 30 percent of cells showed a lap-specific pattern. Some of them were highly active when a rodent ran through the location it responded to on the first lap and relatively quiet during the remaining three laps. Others responded on the second lap far more than the rest, and so on. These neurons, which the researchers termed “event-specific rate remapping,” or ESR, cells, seemed to signal which lap a mouse was on.
To confirm the ESR cells were really encoding events, the researchers conducted experiments using tracks that were elongated along one dimension, increasing their length. Even when lap length was randomly altered between trials, the cells were still much more active on their preferred lap, showing the activity could not be related to the time elapsed or distance travelled. “The results support the idea that the hippocampus can express representations of relevant variables, including, in this case, the number of laps since a reward was delivered,” says Frank, who was not involved in the study.
In another experiment, the team trained mice on a square track on the first day, then substituted a circular track on the next one. Shifting to a new environment resulted in the ESR cells’ spatial responses being completely remapped onto the circular track. Strikingly, though, the lap that those neurons preferentially responded to remained the same. These findings suggest that ESR activity represents segmented units of experience—and that this “event code” can be transferred between different experiences that share a common structure.
Tonegawa compares this process to a familiar scenario. “If you go to a restaurant to have dinner with your friend, that episode is made up of different segments: you arrive at the restaurant, then order an appetizer, then you choose a main dish, and then, usually, you have dessert,” he says. “As all this is going on, the stimuli coming to you are changing. But at the same time, it’s made up with distinct events, where you switch from the appetizer, to eating a main dish, dessert, and so on.” The coding revealed in the study may explain how the brain abstracts events such as “main course” across different visits to different restaurants with different friends. And this idea may offer insight into how the brain generalizes knowledge to learn efficiently. “You’re transferring knowledge you already have, based on past experience, to learn new things,” Tonegawa says. “That’s why we can learn things much faster.” These insights, he thinks, could help engineers develop AI systems with the ability to transfer competencies from one environment to another, such as for medical robots moved between hospitals.
The circular track experiment showed that brain responses that specify your precise location can be altered without affecting event-specific activity. In a final experiment, the team asked whether the reverse is also true. A region called the medial entorhinal cortex (MEC) works closely with the hippocampus in spatial cognition and navigation. There is also evidence that it is involved in segmenting experience into sequential events. The researchers used optogenetics (a technique involving genetically altering cells so they can be activated or inhibited using light) to switch off signals from the MEC to the hippocampus while mice performed the running task. Doing so had no effect on location-specific responses but completely disrupted lap-specific ones, suggesting place and event encoding can be separately manipulated—even though the same cells process both aspects of experience.
One limitation of the study is that running repeatedly around a track is unlike most natural experiences. “There’s no demonstration that these event-related patterns exist the first time an animal experiences a set of events—only that they appear after many repeats of a now familiar sequence,” Frank says. “This is not really the same as our episodic memories, where each new experience gets encoded separately and stored as an event the first (and often only) time it happens.” He thinks the cells represent “well-learned and relevant elements of an experience with repeating elements.” That arrangement, he says, is reminiscent of reports from studies of hippocampal neurons that “fire similarly, but not identically, in geometrically repeated elements of the same environment.”
“This is an insightful experiment, performed with the care and numerous controls characteristic of Tonegawa’s lab,” says neuroscientist György Buzsáki of the NYU Grossman School of Medicine, who was not involved with the study, though he provided comments to the researchers. But Buzsáki has a more radical take on what is happening. He thinks all of the properties researchers have assigned to hippocampal neurons are different aspects of the same fundamental mechanism. To explain this idea, he compares it to the relationship between the motion of a vehicle’s engine and its distance travelled and journey time— different variables reflecting a single underlying process.
In the case of episodic memory, the hypothesized elements are what, where and when. “The definition of episodic memory is: ‘What happened to me, where and when?’” Buzsáki says. When you combine these elements, it re-creates the event. “This is called memory,” he adds. Researchers relate the activity they observe to what, where or when, but all the hippocampus is doing is efficiently encoding experience into a neuronal sequence. The hippocampus is “like a librarian that tells you to go to shelf five, row two. Then the next book is this, then this, and so on,” he says. But the librarian is blind to the content of these sequences, which is constructed in the cortex. Thus, Buzsáki’s interpretation of the new findings is that cells do not encode abstract “event-specific” properties—such as which lap number or dinner course one is experiencing—so much as they generate the ordinal sequences that give memory the order necessary for us to make sense of it.