FriSem

Satchel Grant, Ph.D. Student in the Department of Psychology

Title: Recurrent Models of Language as a Cognitive Technology

Abstract: The interactions between language and cognition are not well understood. How much does language shape human cognition as a learning constraint? To what degree is language used as a cognitive tool in processing specific tasks? How much is language merely a reflection of deeper cognitive processes? In this work we explore

these questions within the realm of numeric cognition. We revisit a debate on the role of exact count words in numeric matching tasks. To address these issues, we introduce virtual environments that simulate the numeric matching tasks used to study the Pirahã in previous works. We then use Long Short-Term Memory networks (LSTMs) to model visuospatial counting behavior with and without the influence of exact number words.

We find that it is possible for networks to learn to perform numeric matching tasks up to non-trivial quantities with and without the use of exact number words. Number words, however, decrease the amount of experience needed to learn the numeric matching

tasks. Furthermore, through latent vector analysis, we find that the policy networks solve the numeric matching tasks using an increment up, decrement down strategy; and lastly, we find evidence of a logarithmically compressed mental number line.