Summary: Researchers have discovered that the mouse brain can hold and process multiple competing hypotheses about spatial location while navigating environments with ambiguous landmarks. In a demanding task where identical cues could indicate different places, neurons in the retrosplenial cortex (RSC) produced distinct activity patterns corresponding to different possible locations. Once additional information resolved the ambiguity, those neural patterns converged on the correct representation, showing the brain actively uses these hypotheses to guide behavior. This is the first direct observation of hypothesis-based navigation coding in the brain.
These neural representations did not merely store alternative possibilities: they were dynamically used to compute the appropriate action to reach the correct goal. The study links biological neural dynamics to artificial networks trained on comparable tasks, suggesting common principles underlying sequential spatial reasoning.
Key Facts:
- Multiple hypotheses encoded: Populations of RSC neurons maintain separate activity states for different candidate locations during ambiguous navigation.
- Decision-making in action: When sensory evidence clarified which landmark referred to the reward, the neural activity collapsed into the representation that led the mouse to the correct port.
- Brain–AI parallel: Patterns observed in the mouse RSC resembled dynamics in recurrent artificial neural networks trained on the same problem, highlighting low-dimensional recurrent dynamics as a possible mechanism.
Source: MIT
Navigating with ambiguous landmarks
When humans or animals navigate environments where landmarks are not unique, they often hold several candidate interpretations in mind and update them as new information arrives. For example, finding an office in a row of similar brick buildings might require remembering that you need the second building on the block rather than relying on distinguishing visual features of any single structure. Similarly, mice navigating an arena with identical or ambiguous cues must use memory of their recent movements and headings to infer which landmark corresponds to a goal.
MIT neuroscientists recorded activity in the retrosplenial cortex, a brain region that integrates visual input, hippocampal spatial signals, and thalamic information to support navigation. Building on prior work showing that RSC neurons combine visual and self-motion cues to mark landmarks, the team designed a more challenging behavioral task to probe whether RSC can form and act on internal hypotheses.
Mice were placed in a circular arena with 16 small openings around the perimeter; only one opening delivered a reward when the mouse poked its nose through. In initial training, a single dot of light on the floor indicated the rewarded port and became visible only when the mouse approached. Once mice learned that cue, the experimenters introduced a second, identical dot. The two dots always sat the same distance apart and from the arena center, but only the counterclockwise dot marked the rewarded port.
Because the lights were identical and visible only at close range, mice could never see both simultaneously and could not instantly distinguish which dot signaled reward. Instead, they had to combine memory of their recent position and heading with evolving sensory input to infer which landmark was which. The task therefore forced animals to form, maintain, and update competing spatial hypotheses until further information disambiguated the situation.
While mice approached the ambiguous landmarks, recordings from RSC revealed that distinct neural population patterns emerged that reflected the different candidate interpretations. Each pattern corresponded to a hypothesis about where the mouse believed the reward lay. As soon as the animal received enough information to identify the correct dot, neural activity shifted and collapsed into the single representation corresponding to the correct location—after which the animal navigated to that port.
These dynamics show that RSC does more than passively hold possible states: it encodes hypotheses as distinct trajectories in population activity space and uses recurrent interactions to update and select among those hypotheses for action. “We show that RSC has the required information for using short-term memory to distinguish ambiguous landmarks, and that these hypotheses are encoded and processed in a way that allows RSC to solve the computation,” said lead researchers.
Interconnected dynamics and artificial networks
When interpreting their results, the researchers compared them with prior work using recurrent artificial neural networks. Those networks, trained on analogous navigation tasks, exhibited activity patterns conceptually similar to the biological data: low-dimensional, highly interconnected dynamics that represent multiple candidate states and evolve to select the correct one. The RSC neurons also showed such structured, low-dimensional dynamics, pointing to interconnectivity as a crucial feature for holding and manipulating multiple hypotheses simultaneously.
Future work will examine how RSC cooperates with other brain regions involved in navigation and decision-making—such as prefrontal cortex and hippocampus—particularly during more naturalistic foraging when animals learn without explicit training. Understanding how these circuits learn and choose what to remember during free behavior could reveal general principles of sequential cognitive learning.
Funding:
This research was funded in part by the National Institutes of Health, a Simons Center for the Social Brain at MIT postdoctoral fellowship, the National Institute of General Medical Sciences, and the Center for Brains, Minds, and Machines at MIT, supported by the National Science Foundation.
About this neuroscience research news
Author: Sarah McDonnell
Source: MIT
Contact: Sarah McDonnell – MIT
Image: The image is credited to Neuroscience News
Original Research (open access): “Spatial reasoning via recurrent neural dynamics in mouse retrosplenial cortex” by Mark Harnett et al., published in Nature Neuroscience.
Abstract
Spatial reasoning via recurrent neural dynamics in mouse retrosplenial cortex
Sensory inputs gain meaning through prior context. Recurrent neural dynamics are thought to interpret stimuli according to such context, but whether they can internally generate and use multiple hypotheses to resolve ambiguity has been unclear. Here, recordings from mouse retrosplenial cortex reveal that RSC can form several evolving hypotheses and perform spatial reasoning through recurrent dynamics. Using ambiguous landmarks that must be identified by their mutual spatial relationships, mice sequentially refined candidate interpretations. Neurons in both biological RSC and trained artificial recurrent networks encoded mixtures of hypotheses, location, and sensory input, constrained by robust low-dimensional dynamics. RSC represented hypotheses as locations in activity space with divergent trajectories for identical sensory inputs, enabling their correct interpretation. These results suggest that interactions between internal hypotheses and external sensory data in recurrent circuits provide a substrate for complex sequential cognitive reasoning.