Mapmaking in the mind: How the brain builds mental maps of the world

Neuroscientists know a lot about the activity of neurons that make up these maps -- like which cells fire when we're in a particular location. But how the brain creates these maps as we learn remains a mystery.

Now, by tracking the activity of thousands of neurons over days and weeks as an animal learns, researchers at HHMI's Janelia Research Campus have systematically detailed, step by step, how these cognitive maps form in the brain's hippocampus -- a region responsible for learning and memory.

The team, led by the Spruston Lab, found that as an animal learns to collect rewards from two subtly different linear tracks -- like the different floors of a hotel -- neurons in the hippocampus start to respond in disparate ways. Eventually, the brain produces entirely distinct representations of these visually similar tracks that include information enabling the animal to differentiate between the two options.

The researchers also identified the type of mathematical model that best reproduces this learning process, shedding light on computations the brain might be using to create these mental maps and providing insight into memory and intelligence.

"We are mapping out the step-by-step process of cognitive map formation, which is such an important concept," says Weinan Sun, an assistant professor at Cornell University who co-led the research as a research scientist in the Spruston Lab. "But there is also a second contribution: The result of watching that process gives us a hint about the underlying computations and we get a little bit closer to understanding what the brain is doing in making these maps."

Understanding how the brain implements these computations could help researchers develop better treatments for memory disorders like Alzheimer's and create artificial intelligence systems that reason more like biological brains.

"Neuroscience and AI can learn a lot from each other," says Johan Winnubst, lead scientist of neuroanatomy at E11 Bio, who co-led the research as a research scientist in the Spruston Lab. "What large language models are able to do is very impressive, but they also fail in a lot of very obvious ways and some of that has to do with reasoning and long-term planning. So maybe you introduce some of the lessons that we have learned from the hippocampus to these models."

Observing map formation

To see how these cognitive maps form, the researchers used a Janelia-designed, high-resolution microscope with a large field of view to image neural activity in thousands of neurons in the hippocampus of a mouse learning how to navigate two different virtual corridors: one with a reward at a near location, and one with a reward at a far location.

Near the beginning of each corridor, the mouse is given a visual cue to indicate where in the corridor it can expect to find a water reward, at either the near or far location. The mouse must figure out the relationship between the indicator cue and where the reward is going to be delivered.

The researchers saw that all the animals learned how to navigate the corridors in the same specific sequence. First, they learned to suppress their licking where they knew they wouldn't be rewarded. Then, they learned they were only getting one reward per corridor. Lastly, they learned to suppress their licking at the near reward location in the corridor where the reward was at the far location.

As the animal started to learn, its neural activity started to reflect its changes in behavior. At the beginning of learning, the activity of individual neurons was mostly similar for the two corridors, forming a linear track with only slight differences representing the different cues and reward locations.

However, as the animal's learning progressed, the neural activity representing the different corridors started to differentiate further. While the near and far reward locations were always represented differently from each other, now these reward locations were treated differently depending on which corridor the mouse was in: The near location in the near corridor was represented differently than the near location in the far corridor, even though they were visually identical.

At the end of learning, the activity of these neurons was completely different, with distinct maps encoding hidden information that enabled the animal to distinguish between the two corridors. The researchers found that there are specific cells -- they call them "state cells" -- that extract hidden information from the environment to enable this differentiation.

In the hotel analogy, initially the brain might represent all the floors similarly. But after a few days, we learn the differences between the floors. Our brains generate different maps for the different floors, each containing hidden or contextual information -- like what number was displayed inside the elevator but is no longer visible when we get out -- that allows us to distinguish between them. For animals in the real world, this process helps to distinguish similar but different areas in a forest or field.

"Initially, the brain activity is very similar, and with learning, the activity becomes more and more different until they are orthogonal. And then, in the end, each neural pattern of activity will encode a hidden state that will reflect the true hidden state of the task," Sun says. "The brain cares about the immediate sensory input but interprets it in the context of the hidden state the animal is in."

Finally, the researchers looked at what computations might be happening in the brain to enable the map formation they observed.

The team discovered that the brain builds these maps like a state machine -- a system that figures out true situations by inferring hidden states beyond what is immediately visible. Among various computational models tested, only one type -- called a Clone-Structured Causal Graph -- could accurately reproduce this learning process.

The researchers, who also created an online visualization tool so scientists around the world can explore the data, say that being able to connect these pieces -- from behavior to individual cells to groups of neurons to algorithms -- is a critical step toward truly understanding how the brain and intelligence works.

"One of the ultimate goals of neuroscience is: if we observe a behavior or cognitive function, we want to understand that behavior or cognitive function in terms of not only the cellular and molecular processes responsible for it, but also the algorithmic representation the brain uses," says Janelia Executive Director Nelson Spruston, the senior author on the new research. "We are getting at the algorithmic level -- arguably the hardest to pin down -- which helps us connect the dots of how the cellular and molecular processes actually operate to produce an algorithm in the brain that can form this computation that we observe in the form of behavior."