Bethia Sun

Hi! My name is Bethia. I’m a second year Computer Science and Engineering MPhil student at UNSW Sydney, advised by Yang Song and Maurice Pagnucco. Previously, I completed a dual B.S at the same institution in Mathematics and Computer Science. I’m currently a visiting scholar at UC Berkeley’s Redwood Center for Theoretical Neuroscience, and the Computation & Language Lab, working with Professors Fritz Sommer and Steve Piantadosi.

My current work centers on understanding how compositional and symbolic structure can emerge within continuous representational systems. I first became interested in this question while working in deep learning, where I grew disillusioned with neurosymbolic approaches that presuppose discrete, predefined symbols. I wanted instead to understand whether compositionality could be realised directly within a distributed representational algebra – whether a vector space itself could encode structured, interpretable composition.

I’ve studied Tensor Product Representations (TPRs) and Vector Symbolic Architectures (VSAs) as mathematical frameworks that make compositional structure explicit in vector spaces while supporting constituent recovery and recursive binding. My current work focuses on developing a formalisation of what it means for a system to be compositional in the abstract algebraic sense, using tools from universal algebra to characterise compositionality independently of any specific implementation or binding operator. In parallel, I’m exploring the geometric and algebraic properties of tensor-product spaces more broadly, as a representational substrate capable of expressing both compositional and non-compositional phenomena ubiquitous in cognition and perception. Feel free to take a look at the interests section of my home page to see how all my interests fit together!

My broader goal is to develop a theory of intelligence that, like physics for motion and matter, bridges scales – explaining how microscopic laws (the continuous dynamics of neural systems) give rise to macroscopic regularities (the discrete, compositional structures of thought). I hope to approach this unification through the mathematical languages of group and representation theory, which formalise invariance and structure; and through inspiration from statistical physics and quantum mechanics, which reveal how complex, emergent behaviour arises from simple local interactions. More broadly, I’m drawn to the idea of a science of intelligence: a theoretical framework that, like physics, seeks deep mathematical principles capable of unifying seemingly strange and diverse phenomena under a common set of laws.

Outside of AI research, I enjoy reading books and watching film, snorkelling, and trying to confirm that the Skinnerian resposnse exists in my feline child.

I’m always happy to discuss a variety of random topics, so feel free to reach out if you’re interested in having a chat!