We build and analyze agent-based models, such as recurrent neural network controllers trained with deep reinforcement learning, to study natural behavior and the neural dynamics that generate it. Like digital twins, these models allow rapid hypothesis testing in silico while staying fully observable. We also develop theory and methods to interpret and compare such networks, and apply the broader machine learning and simulation toolbox to problems in biology and medicine.
The most up-to-date list of our publications is on Google Scholar, sorted newest first.
Building embodied and multi-agent models to study behavior, interaction, sensing, signaling, and adaptation.
Multi-agent RL agents that sense their surroundings and signal to each other with self-generated electric fields, modeled on weakly electric fish. A 3D biomechanical body model is in progress.
RLDM 2025, spotlight talk · NeurIPS 2025 workshop
Deep RL agents with recurrent networks, trained to hunt prey, are used to dissect the hunting behavior of larval zebrafish.
CCN 2026 · NeurIPS 2025 AI4Science workshop, spotlight
How collision avoidance and the distribution of food shape collective foraging in groups of fish agents.
CCN 2026
Insect-like agents trained with deep RL to track turbulent odor plumes. Their behavior resembles that of flying insects, and their recurrent networks learn to compute task-relevant variables with distinct structure in population activity.
Multi-level mechanistic interpretability, before it had a name: the agents were analyzed at the levels of behavior, computation, and neural dynamics together, the way a neuroscientist would study an animal (preprint, 2021). The work is now a canonical NeuroAI example in the Neuromatch NeuroAI course.
Nature Machine Intelligence, 2023 · cover feature
Also: Neuroprospecting with deep RL agents, an essay on what AI and neuroscience can learn from each other · NeurIPS 2021 workshop
Understand what recurrent networks and biological circuits compute, how they implement it dynamically, and why different solutions emerge.
Measuring and controlling how differently task-trained recurrent networks solve the same task.
NeurIPS 2025, spotlight
The benefit of curricula in reinforcement learning depends on reward sparsity and on competing optima.
NeurIPS 2026
As an RL controller learns to balance a pendulum, training prunes its recurrent dynamics into a ring that encodes the pendulum's angle, and stable and unstable fixed points emerge in turn.
RLC 2024 workshop, talk
Learn predictive models from biological and biomedical data, especially when dynamics, interventions, and uncertainty matter.
Most studies of human movement use short, scripted lab tasks. We instead mined weeks of continuous intracranial recordings and video from hospitalized patients for natural, unscripted movements, using pose tracking to find and annotate thousands of arm movements automatically (the clip above, from the J. Neuroscience Methods paper).
J. Neuroscience Methods, 2021 · eNeuro, 2021 · Scientific Data, 2022
Multimodal deep learning with ensemble-based uncertainty to detect junctional ectopic tachycardia in pediatric cardiac ICU patients.
Heart Rhythm O2, 2024
Figure: Ju et al., Heart Rhythm O2 (2024), CC BY-NC-ND 4.0
RACER: an LLM-powered method for scalable, auditable analysis of semi-structured mental health interviews.
EMNLP 2024 NLP4Science workshop
A proposal to decipher electric-fish signals with unsupervised machine translation between signals from multi-agent RL agents and recordings of real fish, constrained by cycle consistency.
NeurIPS 2025 AI for Animal Communication workshop
Using a physics model of the electric field to find where each fish is and which fish produced each signal, so that pose annotations and signal attribution can be automated.
NeurIPS 2025 AI for Animal Communication workshop
Also: binding affinity and gene regulatory network evolution · ALIFE 2023
More papers, code, talks, and helpful links are on Satpreet (Sat) Singh's research page → and Google Scholar.
