Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Dendritic non-linearity supports the formation and reactivation of word memories as cell assemblies
Poster A74 in Poster Session A, Tuesday, October 24, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
Also presenting in Lightning Talks A, Tuesday, October 24, 10:00 - 10:15 am CEST, Auditorium
Alessio Quaresima1, Hartmut Fitz1,2, Renato Duarte2, Peter Hagoort1,2, Karl Magnus Petersson1,3; 1Max Planck Institute for Psycholinguistics, Nijmegen, The Netherlands, 2Donders Institute for Brain, Cognition and Behaviour, Nijmegen, The Netherlands, 3Faculty of Medicine and Biomedical Sciences, University of Algarve, Portugal
During speech comprehension, words are recognized based on the order of phonemes (e.g., /b/æ/t/ versus /t/æ/b/) and retrieved from long-term memory. Behavioral studies demonstrated that lexical access happens in a cascade manner, and items in the lexicon are co-activated (McQueen & Gaskell 2008); however, the underlying neurobiological mechanisms of storage, maintenance, and retrieval are poorly understood (Poeppel & Idsardi 2022). One idea is that word memories are learned through Hebbian plasticity and maintained and reactivated as neuronal assemblies (Buzsáki 2010). To test this idea, we built a neurobiologically-constrained network model of word recognition. The network is composed of spiking neurons with dendrites and we hypothesized that dendritic structure might be crucial to induce and reactivate word assemblies that are structured in time. We simulated two 5-minute-long phases to test the formation and retrieval of word memories. In the association phase, neurons were exposed to phoneme sequences and target words, represented as overlapping cell populations. In the retrieval phase, networks only received phoneme sequences as input, and we measured the reactivation of the corresponding word populations. We tested the network model on seven distinct lexicons of increasing complexity, with 10-15 words each, and varying numbers of phonemes that were shared between words. We compared a network of neurons with dendritic structure (Quaresima et al., 2022) with two networks of point neurons without dendrites (Duarte & Morrison 2019, Litwin-Kumar & Doiron 2014). The dendritic neurons have segregated branches with non-linear, voltage-gated N-methyl-D-aspartate receptors (NMDARs) which endow them with short-term dendritic memory. All networks implemented spike-timing-dependent plasticity on excitatory synapses and were homeostatically balanced with inhibition. When presented with phoneme sequences in the retrieval phase, the dendritic network reactivated the correct word populations with high accuracy (above 70%), including words composed of the same phonemes in a different order (e.g., /d/o/g/ versus /g/o/d/). In contrast, networks of point neurons reactivated only words containing phonemes that were unique to these words, and confused words with shared phonemes (success rate below 20%). Phoneme overlap between words was the main source of difficulty for these networks while lexicon size was less important. In order to identify why the dendritic network was superior, we systematically compared variants of these networks. We found that dendritic memory, supported by slow, non-linear synaptic integration at NMDA receptors, was critically responsible for binding phoneme information over time. The analysis also showed that retrieval accuracy was modulated by the strength of inhibitory control exerted on dendrites. Heterogeneity in integration timescales, induced by asymmetric dendritic lengths, only played a minor role. Our study indicates that word memories can be stored through long-term, associative plasticity and reactivated reliably when dendritic structure is modeled explicitly. Thus, we propose a linking hypothesis between a psychological phenomenon (word recognition) and a neurobiological mechanism (dendritic non-linearity). Future work needs to test whether this mechanism is also able to encode the detailed semantic and morphosyntactic feature structure of words.
Topic Areas: Speech Perception, Computational Approaches