Goal-seeking compresses neural codes for space in the human hippocampus and orbitofrontal cortex

Abstract

Humans can navigate flexibly to meet their goals. Here, we asked how the neural representation of allocentric space is distorted by goal-directed behaviour. Participants navigated an agent to two successive goal locations in a grid world environment comprising four interlinked rooms, with a contextual cue indicating the conditional dependence of one goal location on another. Examining the neural geometry by which room and context were encoded in fMRI signals, we found that map-like representations of the environment emerged in both hippocampus and neocortex. Cognitive maps in hippocampus and orbitofrontal cortices were compressed so that locations cued as goals were coded together in neural state space, and these distortions predicted successful learning. This effect was captured by a computational model in which current and prospective locations are jointly encoded in a place code, providing a theory of how goals warp the neural representation of space in macroscopic neural signals.

Hippocampal sequences span experience relative to rewards

“…These sequences could then act as a substrate that can become anchored to particular salient reference points 16,17,27,47 , allowing maximal flexibility to learn new information as it becomes relevant and layer it onto existing representations [159][160][161] . This role for hippocampal sequences extends to computational studies 28,159,162 and human research [163][164][165][166][167] , where such coding schemes have been linked to the organization of knowledge and the chronological order of events in episodic memory 27,49,[163][164][165][166][168][169][170] . The ability of the hippocampus to parallelize coding schemes for different features of experience may help both build complete memories and generalize knowledge about specific features without interfering with the original memory.…”

Building compositional tasks with shared neural subspaces

“…This may lead to continual learning, which would be facilitated by shared representations 35 . Instead, we found task-irrelevant dimensions were compressed 28,36 as a function of the animals' internal belief about the task state (Fig. 5).…”

Neural representation in active inference: Using generative models to interact with—and understand—the lived world

“…Recent research into the geometries of neuronal responses across various brain regions has gained momentum 75,76,133 . To the extent that these geometries reflect how we perceive the world, it becomes evident that not only the generative model itself but also the specifics of its learning statistics play a crucial role in determining what it represents and how it does so.…”

“…In essence, the most effective generative model tends to retain sufficient detail to accurately represent observations and achieve task success (in the case of action‐guiding models), while discarding redundant information that does not bring an increase in accuracy. Consequently, learning leads to the development of parsimonious models that abstract away from irrelevant details, forming compressed latent spaces or manifolds 60,75–77 . This is evident in the cognitive maps discussed earlier.…”

“…Instead, a greater number of place cells are typically found in locations tied to goals and other significant features 78–80 . Moreover, spatial locations with equal value tend to cluster more closely in the latent or hidden state space, further exemplifying these distortions 76 …”