A central question in neuroscience is how brains generate efficient yet structured representations of the world that are useful for future inferences. Multiple lines of evidence suggests that this structure is compositional, in the sense that the meaning of these representations are composed of and determined by its parts. What neural mechanisms might mediate these processes, and how can we test such theories? In this talk, I suggest that holographic reduced representations, also known as hyperdimensional computing or vector symbolic architectures, provides promising conceptual tools to answer these questions. That is, we can develop both testable neural predictions for how the system does function, and provide functional explanations for why networks of neurons are organized in that way. I will present my own dissertation work on this topic, with a focus on two chapters* about the hippocampus and entorhinal cortex. I will also share my perspective on some of the exciting advances and frontiers in the interdisciplinary landscape of systems neuroscience, theoretical computer science, and cognitive science more broadly. (Keen observers may notice that the talk title itself pays homage to Tony Plate’s dissertation, Distributed Representation and Nested Compositional Structure.)

*They can be found here: “Computing with residue numbers in high-dimensional representation” and “Binding in hippocampal-entorhinal circuits enables compositionality in cognitive maps“.

—

To request the zoom link, send email to jteeters@berkeley.edu.  Also indicate if you would like to be added to the email list for announcements about Redwood Seminars.