Wenrui Ma

Hi, I’m Wenrui

I’m a first-year PhD student in machine learning, advised by Dr. Eva Dyer in the NerDS Lab. I build models that learn transferable structure from messy real-world time series, from neural recordings to wearable sensors.

Before my PhD, I earned an M.S. in Mathematics and an M.S. in Computer Science at Georgia Tech, a B.S. in Computer Science at Colorado State University (Magna Cum Laude), and a B.Eng. in Software Engineering at Hunan University.

A unified view of time series

Foundation models transfer easily across images and text, but time series (the native language of biology and embodied systems) still resist it. I treat neural recordings, wearable sensor streams, robot trajectories, and macro-scale series as different observations of underlying dynamical systems, each measured through some configuration of channels.

My work builds representations that disentangle a system's temporal dynamics from its measurement-channel geometry, so one pretrained model can adapt to new subjects, devices, and channel sets, and serve both classification and forecasting. I work across biological signals (EEG and intracranial recordings, currently cross-subject iEEG in OCD), physiological and robot-trajectory data, and macro-scale financial series.

Papers

figure

ICLR 2026 · TSALM Workshop · Oral

Narrative of Time across Scales (NoTS)

Wenrui Ma, Ran Liu, Ellen L. Zippi, Christopher M. Sandino, Juri Minxha, Behrooz Mahasseni, Erdrin Azemi, Ali Moin, Eva L. Dyer

Reframes time-series pretraining as autoregressively predicting the next scale in a coarse-to-fine decomposition of the signal — a narrative across resolutions rather than across time steps. Lightweight channel and task prompts then adapt one frozen backbone to unseen sensors and to tasks like classification, imputation, and anomaly detection.

figure

Preprint

Multi-Level Transport Alignment for Time-Series Domain Generalization

Wenrui Ma, Zihao Chen, Ran Liu, Sergey A. Shuvaev, Eva L. Dyer

Most domain-generalization methods align time series pair-by-pair; MULTA instead learns a single balanced optimal-transport plan over the whole batch, organizing class and domain structure globally. An alignment-gap reweighting concentrates training on the samples whose learned structure strays furthest from that target, improving cross-subject and cross-dataset transfer.

figure

Preprint

BrainWideBench: A Large-Scale, Multi-Task Benchmark for Neuro-Foundation Models

Alexandre Andre, S. P. Mahato, Vinam Arora, Wenrui Ma, et al.

A common yardstick for brain-wide foundation models: one model pretrains across 126 mice and 423 sessions spanning behavior, spiking activity, and anatomy, then transfers to held-out animals on a battery of decoding tasks. It standardizes how “general-purpose” neural models are trained and compared.

figure

Preprint

SCRYER: A Scalable Framework for Forecasting Neural Population Activity

Zihao Chen, Stephen Kwak, Alexandre Andre, Ian J. Knight, Wenrui Ma, Bijan Pesaran, Sergey A. Shuvaev, Eva L. Dyer

Neural activity is dominated by slow background drift, so forecasters trained on average error tend to miss the sparse spikes that actually carry information. SCRYER’s event-driven contrastive objective targets those transients, while unified neuron embeddings and cross-neuron attention handle populations that change between sessions — producing rare positive transfer across species.

figure

Preprint

Mixtures of Localized Diffusion Processes for Time Series Forecasting

Zihao Chen, Alexandre Andre, Wenrui Ma, Ian J. Knight, Mehdi Azabou, Anqi Wu, Sergey A. Shuvaev, Eva L. Dyer

Running diffusion over an entire forecast is wasteful when most of a signal is easy to predict. MeLD decomposes each signal into learnable frequency bands, scores how uncertain each band is, and applies diffusion only to the uncertain bands while confident bands pass straight through an MLP.

figure

Preprint

PRISM: A Hierarchical Multiscale Approach for Time Series Forecasting

Zihao Chen, Alexandre Andre, Wenrui Ma, Ian J. Knight, Sergey A. Shuvaev, Eva L. Dyer

Splits each signal hierarchically in time and frequency into a bank of bands, then learns selection weights that decide which resolutions matter before recombining them. The result is one forecaster that captures slow trends and fine detail together, instead of committing to a single fixed scale.

figure

NeurIPS 2025 · BrainBodyFM Workshop

A Scalable Self-Supervised Method for Modeling Human Intracranial Recordings During Natural Behavior

S. P. Mahato, Jingyun Xiao, Alexandre Andre, Geeling Chau, Wenrui Ma, Ian J. Knight, et al., Eva L. Dyer

Intracranial datasets are hard to pool because electrode layouts differ from person to person. A Perceiver-based model masks a subset of channels and reconstructs them from the rest via learnable embeddings, learning label-free representations that scale with participants and improve decoding of behavior, speech, and audition.

figure

ICML 2024

Balanced Data, Imbalanced Spectra: Unveiling Class Disparities with Spectral Imbalance

Chiraag Kaushik, Ran Liu, Chi-Heng Lin, Amrit Khera, Matthew Y. Jin, Wenrui Ma, Vidya Muthukumar, Eva L. Dyer

Even when every class has equally many training examples, their learned feature covariances have different eigenspectra — a “spectral imbalance” that sample counts can’t reveal. The paper shows this geometric disparity, not class frequency, is what tracks the accuracy gaps between classes.

Get in touch

I’m always glad to connect. Whether you’d like to discuss ideas in time-series representation learning and biosignals, explore a collaboration, or just ask a question, feel free to reach out.