Shimon Edelman - US grants

DESCRIPTION (provided by applicant): Many studies have demonstrated that the amount of speech children receive is related to their vocabulary and language development: typically, the more speech, the better. In addition to quantity, other characteristics of child-directed speech are positively related to language development. The proposed research examines two of these characteristics: variation sets and contingent timing. Variation sets are clusters of related sentences with partial variation in form, such as "Roll the ball. Can you roll the ball? Let's roll it." Contingent speech occurs when parents talk about an object on which their child's attention is focused. The goal of the proposed interdisciplinary research is to discover how children make use of these characteristics of child-directed speech. What underlying learning mechanisms are responsible for the facilitative effects of variation sets and contingent timing on language acquisition? The answer to this question will help explain the positive relation between the amount of speech directed to children and their language growth. The investigators integrate linguistics, computer science, and developmental psychology to study the contributions of variation sets and contingency in child-directed speech to early language acquisition. Their research will combine naturalistic yet tightly controlled word learning studies in young children with advanced computational modeling to elucidate mechanisms of socially embedded learning of nouns and verbs. The first goal of the research is to assess the effects on noun and verb learning of three properties of child-directed speech: a) the presence of variation sets, b) the time lag between utterances within variation sets, and c) the contingency of variation sets on the child's focus of attention. The second goal is to model the possible learning mechanisms that make use of these three properties. The investigators will use a Spike Timing Dependent Plasticity (STDP) model, which is a biologically plausible variety of the Hebbian learning rule that governs connectivity within neural assemblies. The computational model will receive input structured by the same parameters that facilitate learning in the behavioral experiments. Thus, a population of simulated learners will be created whose language learning performance will be directly comparable to that of human children. Mechanisms identified by STDP modeling will help explain how caregiver behavior assists children in tuning neural assemblies to sequences of speech stimuli. Taken together, the proposed studies will, for the first time, allow connections to be made between social interaction, neural organization, and language learning. By illuminating mechanisms by which infants and children learn from caregivers, the findings could inform interventions for disordered language development and help design more effective approaches to second-language instruction. PUBLIC HEALTH RELEVANCE: By illuminating mechanisms by which infants and children learn language from specific features of caregivers'child-directed speech, the findings could inform interventions for disordered language development. The results could also be used to help parents and early child educators create social environments that foster and support language growth. Eventually, this research could be used to design more effective approaches to second-language instruction.

The brain uses similarity to generalize from the known to the unknown. For example, when a person encounters a new type of fruit and has to decide whether or not it is edible, the person must judge how similar it is to already known edible and inedible items. However, the type of similarity that is taken into account matters. A coconut can look like a rock (visual similarity), but for making a decision about edibility the fact that it hangs from a leafy tree (semantic similarity) is key. With funding from the National Science Foundation, Dr. Rajeev Raizada of Rochester University is investigating how the brain uses similarity to respond adaptively to changing circumstances. With an understanding of how types of similarity, such as visual and semantic similarity, are encoded in the brain, it should be possible to decode them from neural signals. In this project, Dr. Raizada is combining brain imaging with computational modeling and behavioral testing. He is developing novel methods of neural decoding to predict the similarity of brain patterns on the basis of computational models of the stimuli that people are perceiving. In addition, the methods are designed to investigate patterns across different people's brain activations.

The novel computational methods being developed in the project could have significant broader impacts, for example, such techniques underpin brain-computer interfaces that attempt to restore communication to locked-in patients. Moreover, the modeling of semantic similarity in the brain has implications for disorders such as semantic dementia. There are also possible implications for technology. The brain responds flexibly to changing circumstances, but artificial systems, in contrast, are all too often brittle. When confronted with circumstances similar, but not identical, to familiar ones, they break down. Insights into how the brain generalizes from the known to the unknown have the potential to transform our knowledge of how the brain achieves its adaptability, opening up new avenues for endowing artificial systems with similar skills.