Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Multidimensional semantic representations emerge from multi-frequency representational similarity learning
Poster Session C, Thursday, October 1, 10:45 am - 12:45 pm, Wangari Maathai
Saskia L. Frisby1, Christopher R. Cox2, Ajay D. Halai2, Akihiro Shimotake3, Takayuki Kikuchi3, Takeharu Kuneida3,4, Yoshiki Arakawa3, Ryosuke Takahashi3, Akio Ikeda3, Riki Matsumoto3,5, Timothy T. Rogers6,7, Matthew A. Lambon Ralph1; 1University of Cambridge, 2Louisiana State University, 3Kyoto University, 4Ehime University, 5Kobe University, 6University of Wisconsin-Madison, 7University College London
Semantic cognition is our ability to understand objects, words and events and to engage in meaningful action. The cortical semantic system includes a “hub”, centred on the ventral anterior temporal lobes (vATL) bilaterally, that synthesises information across modalities and temporal instances into representations that are multidimensional (robins, wrens and ostriches all have feathers and beaks but differ in size and colour; Cox et al., 2024; Jackson et al., 2021; Lambon Ralph et al., 2017). However, it is currently unknown how these representational feats are achieved via neurophysiological activity within in the vATL. Recent evidence from multivariate decoding of human intracranial grid electrocorticography (Frisby et al., 2026) indicates that animacy can be decoded from all frequency bands from theta to high gamma, but that the best decoding performance is achieved when a wide range of frequencies (4 – 200 Hz) are provided as input to the decoder. However, animacy is a coarse approximation of full semantic structure, and it is not clear whether (1) truly multidimensional information is represented only in one or a few frequency ranges, (2) each dimension is represented independently in its own range, or (3) multidimensional structure emerges only when frequencies from a wide range are considered together. To disambiguate these hypotheses, we extracted power and phase from a large sample of ECoG data (n = 18), collected while participants named line-drawings of animate and inanimate objects. We applied an innovative decoding method, representational similarity learning (RSL; Cox et al., 2024) to predict the coordinates of each item on three semantic dimensions derived from a set of semantic feature norms (Dilkina & Lambon Ralph, 2012) using power or phase from multiple electrodes and frequencies. To assess the first two hypotheses, we trained RSL models on power or phase within a single frequency range - theta (4 – 7 Hz), alpha (8 - 12 Hz), beta (13 – 30 Hz), gamma (30 – 60 Hz) and high gamma (60 – 200 Hz). To assess the third, we trained RSL models on frequencies within a wide range (4 – 200 Hz). RSL models trained on power within just one frequency range could predict only one of the three target semantic dimensions. In other words, we found no evidence for multidimensional representation within any individual frequency range and no evidence that dimensions are represented independently in different frequency ranges. However, models trained on the whole frequency spectrum could predict two of the three dimensions, indicating that multidimensional representations emerge only when multiple frequency ranges are considered together. These results constitute evidence that analysing each frequency of neural activity in isolation may obscure the very information that we seek to discover. Future studies should be open to the possibility that information is represented interdependently across frequencies and electrodes and should select methods capable of detecting that kind of neural code (Frisby et al., 2023).
Topic Areas: Meaning: Lexical Semantics, Computational Approaches