Symbolic Compositional Models

In Module 4, we are seeking to understand how complex data and knowledge are represented, manipulated, and learned. We aim to understand how humans perform abstract tasks, some of which are quintessentially human, such as the creation and use of concepts, logical reasoning, mathematics, and language, and long-range planning over long periods. Separate accounts for some of these abilities exist, but they are generally limited by not including perception and learning in real-world settings as well as not sharing knowledge or representation between abilities. Module 4 integrates capabilities from all modules in order to develop models that carry out these larger tasks, such as models that acquire language and allow agents to follow commands. Within Module 4, our projects span low-level perception for the symbols needed for higher-level reasoning, heading toward how to structure image understanding capabilities, expanding those capabilities to include social understanding and culminate in understanding long plans and stories.

Recent Publications

C. Ross, Berzak, Y., Katz, B., and Barbu, A., Learning Language from Vision., in Workshop on Visually Grounded Interaction and Language (ViGIL) at the Thirty-third Annual Conference on Neural Information Processing Systems (NeurIPS), Vancouver Convention Center, Vancouver, Canada, 2019.
CBMM Funded
Y. - L. Kuo, Katz, B., and Barbu, A., Deep Compositional Robotic Planners that Follow Natural Language Commands., Workshop on Visually Grounded Interaction and Language (ViGIL) at the Thirty-third Annual Conference on Neural Information Processing Systems (NeurIPS). Vancouver Convention Centre, Vancouver, Canada, 2019.
CBMM Funded
C. Han, Mao, J., Gan, C., Tenenabum, J. B., and Wu, J., Visual Concept-Metaconcept Learning, Neural Information Processing Systems (NeurIPS 2019). Vancouver, Canada, 2019.PDF icon 8745-visual-concept-metaconcept-learning.pdf (1.92 MB)
CBMM Funded
CBMM Funded