Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Synergistic Predictive and Reinforcement Mechanisms Drive Language Learning Success
Poster Session C, Saturday, September 13, 11:00 am - 12:30 pm, Field House
This poster is part of the Sandbox Series.
Shuguang Yang1, Suiping Wang1, Gangyi Feng2,3; 1Philosophy and Social Science Laboratory of Reading and Development in Children and Adolescents, South China Normal University, 2Department of Linguistics and Modern Languages, The Chinese University of Hong Kong, 3Brain and Mind Institute, The Chinese University of Hong Kong
Language learning in adulthood presents significant challenges, often resulting in proficiency that is far from native-like and exhibiting considerable individual differences at the population level. We hypothesize that these challenges arise from suboptimal synergization of different learning mechanisms, with varying levels of this synergy accounting for the individual differences in learning success. To test this hypothesis, we examined how predictive coding and reinforcement learning mechanisms work together contribute to language learning by training three transformer-based models to learn an artificial language (Brocanto2) with distinct learning tasks: a) next-word prediction (GPT-G), where the model updates its internal representations by comparing predicted text to real text; b) grammaticality judgments (GPT-C) involving reinforcement learning, where the model is updated based on corrective feedback; and c) an integrated approach combining both training methods (GPT-GC). We collected functional magnetic resonance imaging (fMRI) data from 102 adult participants during an artificial language learning paradigm, tracking neural activity on the first day of learning and measuring behavioral performance over seven training days. All three models learned the artificial language with high accuracy: GPT-G achieved 91% in generating novel and correct sentences, GPT-C reached 100% in grammaticality judgments, and GPT-GC scored 95% for generation and 97% for judgment, surpassing human participants. Representational similarity analysis showed a strong alignment between the models and human brain activity patterns, indicating that these GPT models developed internal representations similar to those of the human learners, even though they were never explicitly instructed to predict words. The prediction-based and reinforcement-based models demonstrated complementary roles in explaining activation patterns in the human brain. GPT-G showed a predominant alignment, while both GPT-G and GPT-C exhibited an alignment that progressed from sensory-motor regions to language regions as their performance improved. Meanwhile, the integrated model offers an additive explanation for the brain activities observed, supporting the hypothesis that synergistic predictive and reinforcement mechanisms are essential for language learning. Additionally, the integrated model and its corresponding neural activation patterns predicted individual learning performance effectively, not only on day 1 but also in the generalization test performance observed on day 7. These findings suggest that human language learning relies on both predictive processing and reinforcement signals. Our results integrate computational modeling and neuroimaging to demonstrate how these learning mechanisms play complementary roles in developing and refining language knowledge. This study provides a new framework for understanding language learning as a synergistic interaction between prediction and reinforcement learning mechanisms, with implications for language education and rehabilitation.
Topic Areas: Language Development/Acquisition, Computational Approaches