Our goal was to test the hypothesis that only humans can recombine mental symbols flexibly using a mental “language of thought”. We tested this idea by designing artificial mini-languages in which the stimuli are simple, non-verbal, and yet demonstrably involve nested structures; then checking whether monkeys and humans can acquire them.
Sequences of spatial locations in humans. Marie Amalric and Stanislas Dehaene designed a geometrical language that can describe, in a compact manner, spatial sequences of locations. Behavioral research proving that it explains human behavior in a spatial working memory task was published by Marie Amalric in PLOS Computational Biology in 2016, at the onset of this project. With Liping Wang, we published the fMRI version of this task in human adults during eye tracking of these sequences (Neuroimage, 2019). Fosca Al Roumi then acquired and analyzed a full set of human MEG data with 20 subjects during this geometrical task (Neuron, 2021), showing how the human brain “factorizes” a sequence into numerical and geometrical primitives, as predicted by our model.
We next tested whether monkeys possess a similar language. Liping Wang and Stanislas Dehaene started with a much simpler task of memorizing and repeating a spatial sequence of 2, 3 or 4 items, and showed that although macaque monkeys do so in a manner strikingly different from human adults and children, without any understanding of the geometry of the sequence (Zhang et al, 2022). Similar research was performed on baboons, in collaboration with Joel Fagot ( Scientific Reports, 2020).
Geometrical representation of static shapes. Mathias Sablé-Meyer extended the language of thought for geometry to the case of static geometrical shapes. We performed many behavioral experiments showing that humans, but not baboons, are sensitive to shape regularity, as determined by the presence of primitives of parallelism, right angle, equal length and symmetry (Sablé-Meyer et al., PNAS 2021). We also designed a more generic language that can account for all of the major shapes that are repeatedly found in all human cultures throughout the world (e.g. lines, circles, spirals, squares, and their combinations) and proved that, again, the language accounts for human but not baboon behavior (Sablé-Meyer et al., Cognitive Science, in revision).
Language for sound sequences. We designed and tested another mini-language for binary auditory sequences made of two sounds x and Y). The language captures why humans find some sequences regular. We first proved that human working memory for sequences is well captured by this language (Planton et al., PLOS Biology, 2021), then showed that the same is true of human MEG and fMRI signals. Maxime Maheu proved that humans indeed possess two distinct systems for discrete rules (unique to humans) and for statistical regularities (shared with other primates) (Maheu et al., Elife 2019, and Nature human behaviour, 2022).
Representation of natural language in humans. We analyzed the networks for natural language with a similar approach. Yair Lakretz investigated whether and how artificial neural networks implement linguistic structures (Lakretz et al., Cognition, 2021). To experimentally study the neural encoding of sentences, we established collaborations that gave us access to intracranial recordings in the human brain. In collaboration with Prof. Itzhak Fried from UCLA, we now have a database of human intracranial signals, including single neuron firing, during sentence processing, which we believe to be quite unique at the international level (Lakretz et al., in preparation). Other collaborations were launched with Prof. Nitin Tandon from UTHealth center in Houston, Prof. Florian Mormann at the University of Bonn Medical Center, and with Professors Fabrice Bartolomei and Christian Bénar at Aix-Marseille University. Intracranial signals were recorded while patients performed various versions of our previous sentence paradigm (Nelson, et. al, PNAS, 2017). The resultsconfirm that sentence processing in humans involves the construction of recursive tree structures, not just ordinal sequences, and that ramping brain signals are a marker of such construction.
Overall, we proved that (1) the human brain spontaneously compresses linguistic, auditory and geometrical information using recursive tree structures; and (2) monkeys and baboons fail to do so on the very same tests. The results have strong implications for our understanding of the origins of human singularity (Dehaene et al., Trends in Cognitive Science, 2022, in revision).
