Dean Light

I’m an nth-year PhD student* in Yulia Tsvetkov’s lab in the UW NLP group, at the Paul G. Allen School of Computer Science & Engineering, University of Washington.

* n = 1

Before UW

I did an M.Sc. in Biology (cum laude) and a dual B.Sc. in Mathematics and Computer Science at the Technion — Israel Institute of Technology. After my M.Sc. I co-founded two small startups — BiotaX (machine learning for noninvasive cancer detection) and Stealth (a no-code platform for declaratively tuning LLM agents) — In parallel, I worked as a research consultant at the Technion Data & Knowledge Lab, where I focused on finding and incubating new research directions and connecting the lab’s theoretical work to applied use cases.

Dean Light on a summit above the Cascades

Research

My main research is focused on enabling other AI researchers to build safe, inspectable and controllable systems through two complementary approaches in neuro-symbolic AI.

  • Top-down — controllable and inspectable LLM-based agents. Instead of hardcoding or post-training an agent’s reasoning structure offline, we can build agents who meta-reason about their harness logic just-in-time — mimicking how humans change the structure of their thinking per task and throughout task progression. In Deep Reasoning we created a formal language for translating human metacognition into a formal structured representation that can be used by agents. We are now working on making Deep Reasoning systems automatically adapt to new domains based on both environmental and natural-language feedback.

  • Bottom-up — learning directly over structured representations. In modern deep learning, we tend to learn over data by flattening it to a structure-less representation, trying to push our domain understanding to the network through data mixing and augmentations. We are unable to insert our domain expertise into models in a data-efficient way. Structured and geometric deep learning promise to alleviate this burden and let us use discrete logic and differentiable operators to model our domain expertise directly into the network architecture. However, building such neural networks is prohibitively difficult. In the RelaNN project, we introduce a neuro-relational algebra that combines both relational algebra and differentiable algebra, and develop RelaNN, an embedded declarative programming language that tries to make writing complex structured neural networks as simple as writing SQL queries.

What I care about

Beyond the technical questions, I’m drawn to the social stakes: how we navigate online information landscapes, and what neuro-symbolic tools might do about echo chambers, information overload, and the erosion of a shared sense of truth.

Interests: databases & declarative query languages · natural language processing · machine learning · programming languages · information extraction · neuro-symbolic AI · knowledge representation & reasoning · geometric / graph deep learning · software architecture.

See my publications (via Google Scholar) and selected projects.