Export
Research Overview
Michael C. Frank focuses on computational models of language acquisition, pragmatic reasoning, and human visual attention. His trajectory spans from early Bayesian frameworks for cross-situational word learning to probabilistic models of multiple object tracking. The most recent work in his profile investigates how artificial agents and humans learn language through recursive pragmatic inference.
Research Trajectory
Frank’s early work addresses the cognitive mechanisms underlying infant language acquisition, particularly the challenge of mapping words to their referents in ambiguous contexts. He introduced A Bayesian Framework for Cross-Situational Word-Learning, a computational model that leverages both cross-situational observation and social context to infer intended referents. Specifically, the work formalized the "chicken-and-egg" problem of early word learning: whether to rely on the co-occurrence of words and objects across different situations or to use social cues to determine reference. By testing the model on a corpus of mother-infant interactions, this research demonstrated how Bayesian inference can account for well-documented experimental phenomena such as mutual exclusivity, fast-mapping, and generalization. This work contributed to the broader research direction of Resolving Referential Uncertainty in Grounded Language Acquisition and touched upon areas such as Plan Based Models of Intentional Communication.
Expanding his focus to human visual attention, Frank participated in research examining the cognitive architecture of multiple object tracking. In Explaining human multiple object tracking as resource-constrained approximate inference in a dynamic probabilistic model, he and his collaborators utilized a computational analysis to distinguish between human performance failures caused by inherent cognitive limitations and those resulting from inevitable perceptual uncertainty. By developing an ideal observer model, the study demonstrated that many behavioral phenomena previously attributed to specialized cognitive modules or structural limits are naturally produced by the operation of resource-constrained approximate inference. This study advanced the research area of Mapping Computational Principles to Biological Neural Architectures and utilized techniques related to the direction of Rao-Blackwellized Particle Filters for Hybrid State Estimation.
Returning to the domain of language, Frank's subsequent research sought to unify pragmatic inference with word learning. He oversaw the development of Learning and using language via recursive pragmatic reasoning about other agents, which proposed a model where language learners assume they are jointly approximating a shared, external lexicon. While previous work had independently characterized pragmatic inference and word learning in probabilistic terms, this research provided a unified framework. By reasoning recursively about the intentions of other agents, the model captures how language users invent new terms in conversation and how learners acquire literal meanings based on pragmatic cues. This contribution aligns with the research direction of Language Games for Decentralized Semantic Coordination and furthered his ongoing interest in the area of Joint Inference for Circular Dependencies in Language Learning.
Current Directions
The most recent paper in the provided evidence is from 2013, focusing on the intersection of pragmatic reasoning and vocabulary acquisition. Current directions indicated by this work center on recursive probabilistic models of communication, where agents learn language by modeling the intentional states of their conversational partners. This involves formalizing how shared lexicons are negotiated and approximated in context, bridging the gap between short-term pragmatic inference and long-term language learning.
Profile updated
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake