A summary of recent research projects at the intersection of biology, learning, and computation.

In-context learning and associative memory

In-context learning refers to the ability of models to generalize from examples presented only at inference time — without parameter updates. We study this behavior through the lens of associative memory, where a system retrieves stored patterns based on partial input.

ICML 2025 - Selected for oral presentation
We introduce in-context denoising as a probe task to analyze how attention heads process information. We connect this behavior to associative memory (Hopfield networks and their modern, dense variants), with the trained attention mechanism acting as a single gradient step in a context-dependent associative memory landscape.

NeurIPS 2026 (to appear)
We extend this picture to deeper attention-only transformers in which every token is corrupted. We find a context- and depth-dependent associative memory landscape governing inference: depth refines the context through particle dynamics. This two-stage empirical Bayes view gives a distinct statistical role to attention residuals, which preserve the noisy input as a query for posterior inference — the initial condition for a single gradient step on the refined landscape.

Habituation and transient computation

Habituation is a basic form of non-associative learning that enables organisms to suppress repetitive stimuli, yet it lacks a broadly accepted mathematical framework.

We characterize this phenomenon and its hallmarks mathematically, and identify the simplest dynamical circuits (state-space models) that can produce it (PNAS, 2024). We show that such a motif can be built as a simple analog electrical circuit (CDC, 2024). Bringing these models to experiments, we find that reactivity and habituation vary widely across genetically identical flies yet can remain stable for weeks within an individual (PNAS, 2026), and that thin films of a quantum material habituate to repeated electromagnetic stimuli (Advanced Science, 2026).

Annual Review of Control, Robotics, and Autonomous Systems (in press) - Invited review
We study habituation as a form of transient computation from a behavioral (input–output) perspective, and connect it to broader topics in control and machine learning.

Growing networks of coupled oscillators

In biological systems, cellular division is governed by internal oscillations and modulated by interactions with neighboring cells. Motivated by this process, we propose a new class of growing dynamical systems where individual oscillators replicate upon completing a cycle, recursively shaping the collective network.

Development, 2023
We introduce a minimal model in which coupled relaxation oscillators divide into two with each full cycle, generating structured networks of interacting oscillators. Despite having only three tunable parameters, the system reproduces the majority of germline cyst structures observed across invertebrates, offering a dynamical systems perspective on multicellular self-organization.

Collective gene-expression patterns in multicellular systems

How do multicellular systems, which are made up of genetically identical cells, achieve and maintain diverse patterned states while remaining responsive to intercellular and environmental signals?

Cell Reports Physical Science, 2023
We develop a model of interacting cell types based on gene regulatory dynamics and intercellular communication. Cells are represented as associative memory networks, and tissues as graphs of identical cells. The model reveals self-organized behaviors — including multi-stability of tissue-level attractors — emerging from internal gene regulation and cell-cell interactions.

Connecting Hopfield networks to Restricted Boltzmann Machines

Hopfield networks and Restricted Boltzmann Machines are two classic energy-based models, one built to store memories and the other to learn the structure of data.

ICLR 2021 - Selected for oral presentation
We formally show that Hopfield networks with correlated patterns can be mapped to Restricted Boltzmann Machines (RBMs) with orthogonal weights. Experiments on the MNIST dataset demonstrate that this mapping provides a useful deterministic initialization to the RBM weights. Furthermore, the connection to interpretable Hopfield networks offers a lens for understanding the learning dynamics of RBMs.