Symposium in Memory of Jeff Elman: Language Learning, Prediction, and Temporal Dynamics

Jeffrey Locke Elman (1948-2018) devoted his career to studying human language. He investigated how people use language flexibly and productively, and how these abilities are learned from linguistic and other input. Jeff was a faculty member at the University of California, San Diego from 1977 until he passed away in 2018. His early research focused on phonetics and phonology. This work began his theoretical journey that resulted in the ideas for which he is best known: seemingly discrete combinatorial units of language, such as phonemes, may best be understood as emergent properties of underlying continuous multidimensional representations, such as phonetic input. In the early 1980's, Jeff was a key part of transformative developments at UCSD in connectionist modeling, working with Jay McClelland, Dave Rumelhart, Geoff Hinton, and Liz Bates. With McClelland, he developed the TRACE model of speech perception. In TRACE, speech perception is seen as a constraint satisfaction process in which prior and subsequent context combine with incoming sensory evidence to determine how humans perceive speech. Jeff then turned his attention to a central but often neglected aspect of cognition: time. Jeff's work on Simple Recurrent Networks, beginning with his classic 1990 article Finding Structure in Time, proposed that time-evolving continuous hidden-state representations are fundamental to language processing, and enable prediction-based learning of language. This work remains among the most influential in the history of Cognitive Science. Jeff's subsequent work explored the initial conditions under which a simple recurrent network would recover grammatical structure. He then led a collaborative project to rethink the nature of what must be built in as a foundation for language, and more generally for cognition (Elman et al., 1996). In later work, he focused on the relationship between language and event knowledge. He argued that words do not have meanings, but instead provide clues that a listener uses to understand language. He also focused on event knowledge as a basis for prediction during language comprehension (Elman, 2009; Metusalem et al., 2012). Jeff's final major contribution was a model of how event knowledge is learned. He argued that knowledge of the components and temporal structure of events emerges as a consequence of prediction-based learning (Elman & McRae, 2019). Jeff also played a major role in advancing Cognitive Science as a field. At UCSD, he and colleagues co-founded the interdisciplinary Center for Research in Language in 1985. In 1986, Jeff was a major part of the first Cognitive Science department, which he chaired from 1995 to 1998. Jeff also served as Dean of Social Sciences, and a founder of both the Kavli Institute for Mind and Brain and the Halicioğlu Data Sciences Institute. Finally, Jeff provided guidance for the field by serving as President of the Cognitive Science Society, and a highly respected Chair of the NIH Language and Communication study section. This symposium honors Jeff’s memory. The introduction and discussion will be led by the organizers (McClelland & McRae). In between, four speakers whose work reflects the legacy of Jeff’s contributions will present research from the perspectives of cognitive neuroscience, cognition and perception, language development, computational modeling, and deep learning in simulated embodied agents.