Developmental AI for Language and Cognition

Talk title

Are computational cognitive models of human language still relevant?

Abstract

Computational modeling of human cognition has a long history, and as a methodology offers many advantages over traditional empirical methods. Computational models of human language acquisition and processing in particular have been used extensively to simulate the emergence of abstract linguistic knowledge from child-directed input, and to test language learnability hypotheses as well as to explain developmental patterns observed in empirical data.

The rapid rise of deep learning architectures and especially LLMs has been seen as a breakthrough in terms of simulating human-like linguistic behaviour. Many studies now treat LLMs as a proxy for the human brain, and compare their hidden activations with those measured in empirical or neuroscientific studies of human subjects. However, these models are typically developed without conforming to cognitive plausibility constraints such as the quantity and quality of their training data as well as their memory and processing limitations. This disconnect questions the employment of LLMs as valid tools for studying human language and processing from a cognitive perspective.

In this talk, I will review the fundamentals of developing and using cognitive computational models for studying human language, discuss a few alternative approaches to adapting the state-of-art language technology for studying human language, and suggest directions for developing ecologically valid and cognitively informed computational models.

Bio

Afra Alishahi is a Professor of Computational Linguistics at the Department of Computational Cognitive Science, Tilburg University. Her research interests revolve around building and using computational models of human language acquisition. In recent years, her research has focused on grounded language learning from audio-visual signals, learning through interaction, and analyzing and interpreting emerging linguistic knowledge in deep neural networks.