Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
What's in a word? Raw statistical learning sequences emulate neural entrainment
There is a Poster PDF for this presentation, but you must be a current member or registered to attend SNL 2023 to view it. Please go to your Account Home page to register.
Poster C90 in Poster Session C, Wednesday, October 25, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
Qihui Xu1, Guro Stensby Sjuls2, Marina Kalashnikova1,3, James S. Magnuson1,3,4; 1BCBL: Basque Center on Cognition, Brain and Language, 2NTNU: Norwegian University of Science and Technology, 3Ikerbasque, Basque Foundation for Science, 4University of Connecticut
Infants develop sensitivity to wordlike patterns in syllable sequences based on statistical regularities (Saffran et al., 1996). Typically, this is measured behaviorally (e.g., looking time paradigms). Recently, Choi et al. (2020; see also Batterink & Paller, 2017, 2019, 2020) reported converging EEG evidence. They presented syllables at approximately 3hz, making the word rate 1hz. Initially, subjects exhibited neural entrainment phase-locked to syllable rate (3hz, indexed by intertrial coherence or ITC). With exposure, phase-locking at word rate also emerged, potentially reflecting statistical learning, especially since word rate ITC correlates with individual behavioral learning measures. We simulated this paradigm with Simple Recurrent Networks (SRNs) trained on two regimens. The first was based on Saffran et al. (1996), with four trisyllabic pseudowords (labeled ABC, DEF, GHI, JKL). Words were randomly ordered, but words could not immediately repeat. Transitional probabilities (TPs) were 1.0 within words and 0.333 between words. The second was 6 bisyllabic pseudowords (AD, AE, BF, BG, CH, CI) based on the "box language" of French et al. (2011). All TPs were 0.5 because A-words were followed either by B or C, B-words by A or C, and C-words by A or B. We implemented SRNs with 12 input and output nodes (1 per syllable), and 12 hidden and context nodes. SRNs were trained to activate the next syllable given the current one. To simulate time series with syllable rates of 3hz (Saffran) or 2hz ('box'), we used Frank & Yang's (2018) procedure to convert hidden and output states to extended, noisy time series with 3 (or 2) syllables and 1 word per second. We calculated ITC using a conventional method for EEG. SRNs learned both regimens, and simulated human-like preferences to trisyllabic pseudowords vs. "part words" (last syllable of one word and initial one or two syllables of another). Analyses of hidden and output activations revealed human-like patterns, with high ITC at syllable rates (3hz or 2hz) early in learning, and elevated ITC at word rate (1hz) later. This seems to suggest SRNs are good candidate models for SL, since they simulate both behavioral and neural responses. However, we then estimated ITC from raw inputs. Surprisingly, ITC was elevated at both syllable and word rates. We speculate that this is due to subsets of syllables occurring at 1hz intervals (e.g., for Saffran, A, D, G, and J occur in positions 1, 4, 7, etc.; for 'box', A, B, and C occur only in odd positions). This regularity could drive word rate ITC without learning. Control analyses with syllables scrambled only showed high ITC at syllable rates. A system that developed a distinct response to each syllable could exhibit elevated ITC at word rates without learning words. However, ITC still appears to reflect learning (since ITC and behavioral measures correlated in Choi et al.). Our findings suggest that neural entrainment results must be interpreted with caution, and should be paired with converging behavioral evidence. Also, SRNs provide a promising candidate mechanism for explaining SL, though the same caveats apply.
Topic Areas: Computational Approaches, Language Production