2018 Friday Poster 6609

Friday, November 2, 2018 | Poster Session I, Metcalf Small | 3pm

Statistical learning at noisy environment is associated with vocabulary
V. Kozloff, A. Nguyen, J. Arciuli, Z. Qi

Children and adults have the remarkable statistical learning (SL) ability to rapidly detect and extract probabilistic information. Robust SL has been reliably demonstrated in previous studies where subjects were exposed to a continuous stream of stimuli comprising sequences with high internal transitional probabilities (e.g. Saffran et al., 1996; Turk-Browne et al., 2005). In reality, learning often encounters frequent interruptions by random noise. In the current experiment, we examine SL when structured sequences are interleaved with random sequences. We ask whether adults are able to learn statistical information embedded in a noisy environment, and if so whether adults learn better when the structured and random sequences contain stimuli from different domains (linguistic vs. non-linguistic) or the same domain. We also examined the relationship between statistical learning outcomes and individuals’ vocabulary.

Forty-eight adults, randomly assigned to the Same and Different Groups, completed two visual SL tasks in both linguistic (Letter) and non-linguistic (Image) domains (Table 1). The two groups were matched for age, gender, and NIH Toolbox Picture Vocabulary scores (p’s > 0.25). The structured stream comprised four triplets of stimuli, each repeated 24 times. The random stream contained a different set of 12 stimuli ordered randomly. Each stream was equally divided into 6 mini-blocks. The mini-blocks were interspersed between structured and random conditions to form a continuous stream. During the exposure phase, subjects responded to a target stimulus in each mini-block. Immediately following the exposure phase, subjects completed a two- alternative forced-choice task (Fig. 1). We measured both the reaction time (RT) slope during the exposure phase and the response accuracy during the test phase.

Both groups performed significantly above chance during both tasks’ test phases (Fig. 2, p’s < 0.05). An ANOVA revealed no significant effect of group (Same vs. Different, F (1,46) = 1.87, p = 0.18) or task (Letter vs. Image, F (1,46) = 0.50, p = 0.48). The RT slope during the exposure phase was significantly more negative (quicker acceleration) in the Structured than in the Random condition (F (1,44) = 8.88; p = 0.005) and in the Image than the Letter task (Fig. 3, F (1,44) = 5.61; p = 0.02). There was no significant effect involving group. We extracted each learner’s difference score of RT slope (Structured minus Random) as the online learning measure and response accuracy as the offline learning measure for both tasks. Within the Different group, higher vocabulary score was significantly and more strongly associated with the online measure of the Letter task (r = -0.54, p = 0.007) than with the Image task (t = 2.14, p = 0.02, Fig. 4). No significant correlation was found in the Same group or with the offline measures.

In summary, these results suggest adults are capable of learning statistical information even when the useful information is scattered and embedded in noisy environment. Similar task performance between two groups indicate linguistic and non-linguistic SL share core cognitive resources. However, linguistic SL explains more variability in adults’ vocabulary than non- linguistic SL, hinting linguistic-specific constraints for language learning.