https://www.sciencedirect.com/science/article/pii/S1053811921001531
I just read the paper, some interesting points on testing cognitive maps with language and seeing if grid codes will fire.
Their experiment consisted of creating audiovisual objects (words, shapes, and associated sounds) and had them mapped onto a 2D space.
31 participants were recruited and trained on 4 tasks and then finally tested on semantic comparison where a first word was shown, then a second word was shown and then they were asked on of two questions: has there been an increase,decrease or no change in size?” or “has there been an increase, decrease, or no change in pitch?”
They did find grid codes firing in the MEC when doing their experiment.
“As a linguistic species, we humans use arbitrary symbols to express / recover / evaluate concepts without having to experience them directly through the senses: we can decide that a saxophone is louder than a flute without the need to hear their sounds but relying on the long term representation of the associated words. How does the human brain represent this relational information between concepts when they are presented linguistically in the form of words?”
“One possibility is that we store in memory all the possible pairwise relations in a piecemeal format: flute is cheaper than clarinet (flute < clarinet); clarinet is less loud than saxophone (clarinet < saxophone); and so on. Another possibility is that we summarize this information in a single structure that spans multiple dimensions (loudness and price, in the example above) and that serves as a general and comprehensive model of the mutual relations existing between concepts. This latter representational model, evidently more parsimonious, is what Tolman named “cognitive map”, which he thought as a general model for human cognition, but demonstrated only in the domain of spatial memory and navigation (e.g., Tolman 1948).”
“On a related note, our findings also extend the result of Solomon et al. (2019), who used intracranial recordings to demonstrate that neurons in the human hippocampus are sensitive to the semantic distance (as computed using word2vec-derived subspaces) between real words recalled from a list.”
Combining word2vec with grid cells sounds like a fun dream project.
” Do brain regions such as the entorhinal cortex, the hippocampus and the medial prefrontal cortex represent multiple dimensions at all, or they reduce the dimensionality of representations for instance via selection or compression, as few recent studies seem to suggest (e.g., Theves et al., 2020; Mack et al., 2019; Bottini and Doeller 2020)?”
“With the current experiment we showed that two brain signatures typical of cognitive maps for bi-dimensional spaces, the grid-like code and the distance code, are recruited when humans process words – the building blocks of our language – when their relations can be captured by a 2D structure. We also showed that together with such 2D representations, the brain concurrently represents individual dimensions of the word space. This suggests that the brain generates multiple maps of different dimensionalities of the same conceptual space. Several crucial issues remain open, and will direct future research. First, grid-cells in rats have been shown to alter the typical periodicity of their firing rate when the geometrical structure of the physical environment deviates from regular shapes (e.g., rectangular or circular arenas compared to trapezoidal ones, see Krupic et al., 2015), and this is also related with poorer spatial memory (Bellmund et al., 2020). Conceptual spaces are usually more chaotic and less well structured than what we have tested so far in artificial situations and we might expect the grid-like or the distance signatures of a cognitive map to be at least partially degraded when the structure of a conceptual space deviates from perfect regularity and homogeneity. Second, the knowledge we have about things in the world is tremendously rich and usually hardly reducible to two dimensions. Whether these representational codes, or the bi-dimensional representational format usually assumed for the cognitive map, also hold for multidimensional spaces will be an important matter of future investigation. At the moment, we still lack a precise and conclusive description of how the grid-like code might behave when more than two dimensions (that is, three dimensions) are considered during spatial navigation (but see Kim et al. 2017; Kim andMaguire 2018), and this problem is even more pronounced in the case of conceptual spaces.”