About

Dr. Thomas Serre is a Professor in Cognitive, Linguistic & Psychological Sciences and in Computer Science at Brown University. He is also an affiliate of the Carney Institute for Brain Science and serves as the Faculty Director of the Center for Computation & Visualization, as well as the Associate Director of the Center for Computational Brain Science. Additionally, he holds an International Chair in Artificial Intelligence within the ANR-3IA Artificial and Natural Intelligence Toulouse Institute in France. Dr. Serre earned his Ph.D. in Neuroscience from MIT in 2006 and an MSc in Electrical Engineering and Computer Science from Télécom Bretagne in France in 2000. His research focuses on understanding the neural computations that support visual perception, working at the intersection of biological and artificial vision.

Research topics

• Artificial Intelligence
• Computer Science
• Machine Learning
• Engineering
• Computer vision
• Psychology
• Neuroscience
• Control engineering
• Human–computer interaction

Selected publications

• Perilesional neuromodulation replaces lost sensorimotor function in persons with spinal cord injury

  Nature Biomedical Engineering · 2026-03-11

  article
  PublisherDOI

• MTSE: Multi-Target Speaker Extraction for Conversation Scenarios

  2025-08-17

  article1st authorCorresponding
  PublisherDOI

• Performance decreases for untrained orientation observed in dominant computational models but not humans are mitigated by divisive normalization in encoding processes of visual perceptual learning

  Journal of Vision · 2025-07-15

  articleOpen access

  Visual perceptual learning (VPL) refers to long-term performance improvements following visual experience. It is controversial whether VPL arises from plasticity at the level of neural encoding processes or downstream readout processes (Watanabe & Sasaki, 2015, Ann. Rev. Psych.; Dosher & Lu, Perceptual Learning, MIT Press, 2020). To address this issue, we compared how training on an orientation detection task alters performance in humans and in a well-established neural network model of the early visual cortex, which includes broad excitation, untuned inhibition, and readout components. We focused on performance changes in untrained orientations, which have been discussed only relative to the trained orientation. In the psychophysical experiment with humans, as predicted, the greatest performance enhancement was observed at the trained orientation. Enhancements tapered off as orientations deviated further from the trained one, until around 90 degrees, where no performance increase was observed. Performance at orientations approximately 90 degrees from the trained orientation remained unchanged. In the neural network model simulation, plasticity in the readout components always led to unexpected performance decreases at untrained orientations. However, the introduction of plasticity in untuned inhibition during neural encoding processes, leading to a divisive normalization effect, mitigated the performance decreases at untrained orientations. These findings suggest that divisive normalization plasticity may resolve the discrepancy between the results from human psychophysics and the initial computational model. Our results further suggest the involvement of neural encoding processes in VPL.

  PublisherDOI

• From prediction to understanding: Will AI foundation models transform brain science?

  Neuron · 2025-10-22 · 4 citations

  article1st authorCorresponding
  PublisherDOI

• GASPnet: Global Agreement to Synchronize Phases

  ArXiv.org · 2025-07-22

  preprintOpen access

  In recent years, Transformer architectures have revolutionized most fields of artificial intelligence, relying on an attentional mechanism based on the agreement between keys and queries to select and route information in the network. In previous work, we introduced a novel, brain-inspired architecture that leverages a similar implementation to achieve a global 'routing by agreement' mechanism. Such a system modulates the network's activity by matching each neuron's key with a single global query, pooled across the entire network. Acting as a global attentional system, this mechanism improves noise robustness over baseline levels but is insufficient for multi-classification tasks. Here, we improve on this work by proposing a novel mechanism that combines aspects of the Transformer attentional operations with a compelling neuroscience theory, namely, binding by synchrony. This theory proposes that the brain binds together features by synchronizing the temporal activity of neurons encoding those features. This allows the binding of features from the same object while efficiently disentangling those from distinct objects. We drew inspiration from this theory and incorporated angular phases into all layers of a convolutional network. After achieving phase alignment via Kuramoto dynamics, we use this approach to enhance operations between neurons with similar phases and suppresses those with opposite phases. We test the benefits of this mechanism on two datasets: one composed of pairs of digits and one composed of a combination of an MNIST item superimposed on a CIFAR-10 image. Our results reveal better accuracy than CNN networks, proving more robust to noise and with better generalization abilities. Overall, we propose a novel mechanism that addresses the visual binding problem in neural networks by leveraging the synergy between neuroscience and machine learning.

  PublisherOA PDFDOI

• Follow the Energy, Find the Path: Riemannian Metrics from Energy-Based Models

  ArXiv.org · 2025-05-23

  preprintOpen access

  What is the shortest path between two data points lying in a high-dimensional space? While the answer is trivial in Euclidean geometry, it becomes significantly more complex when the data lies on a curved manifold -- requiring a Riemannian metric to describe the space's local curvature. Estimating such a metric, however, remains a major challenge in high dimensions. In this work, we propose a method for deriving Riemannian metrics directly from pretrained Energy-Based Models (EBMs) -- a class of generative models that assign low energy to high-density regions. These metrics define spatially varying distances, enabling the computation of geodesics -- shortest paths that follow the data manifold's intrinsic geometry. We introduce two novel metrics derived from EBMs and show that they produce geodesics that remain closer to the data manifold and exhibit lower curvature distortion, as measured by alignment with ground-truth trajectories. We evaluate our approach on increasingly complex datasets: synthetic datasets with known data density, rotated character images with interpretable geometry, and high-resolution natural images embedded in a pretrained VAE latent space. Our results show that EBM-derived metrics consistently outperform established baselines, especially in high-dimensional settings. Our work is the first to derive Riemannian metrics from EBMs, enabling data-aware geodesics and unlocking scalable, geometry-driven learning for generative modeling and simulation.

  PublisherOA PDFDOI

• From Prediction to Understanding: Will AI Foundation Models Transform Brain Science?

  ArXiv.org · 2025-09-21

  preprintOpen access1st authorCorresponding

  Generative pretraining (the "GPT" in ChatGPT) enables language models to learn from vast amounts of internet text without human supervision. This approach has driven breakthroughs across AI by allowing deep neural networks to learn from massive, unstructured datasets. We use the term foundation models to refer to large pretrained systems that can be adapted to a wide range of tasks within and across domains, and these models are increasingly applied beyond language to the brain sciences. These models achieve strong predictive accuracy, raising hopes that they might illuminate computational principles. But predictive success alone does not guarantee scientific understanding. Here, we outline how foundation models can be productively integrated into the brain sciences, highlighting both their promise and their limitations. The central challenge is to move from prediction to explanation: linking model computations to mechanisms underlying neural activity and cognition.

  PublisherOA PDFDOI

• Is This Just Fantasy? Language Model Representations Reflect Human Judgments of Event Plausibility

  ArXiv.org · 2025-07-16

  preprintOpen access

  Language models (LMs) are used for a diverse range of tasks, from question answering to writing fantastical stories. In order to reliably accomplish these tasks, LMs must be able to discern the modal category of a sentence (i.e., whether it describes something that is possible, impossible, completely nonsensical, etc.). However, recent studies have called into question the ability of LMs to categorize sentences according to modality (Michaelov et al., 2025; Kauf et al., 2023). In this work, we identify linear representations that discriminate between modal categories within a variety of LMs, or modal difference vectors. Analysis of modal difference vectors reveals that LMs have access to more reliable modal categorization judgments than previously reported. Furthermore, we find that modal difference vectors emerge in a consistent order as models become more competent (i.e., through training steps, layers, and parameter count). Notably, we find that modal difference vectors identified within LM activations can be used to model fine-grained human categorization behavior. This potentially provides a novel view into how human participants distinguish between modal categories, which we explore by correlating projections along modal difference vectors with human participants' ratings of interpretable features. In summary, we derive new insights into LM modal categorization using techniques from mechanistic interpretability, with the potential to inform our understanding of modal categorization in humans.

  PublisherOA PDFDOI

• Comparing a Dual-stream Architecture with Single-stream CNNs to Simulate Vision in Locomotor Control

  Journal of Vision · 2025-07-15

  articleOpen access

  Zhu and Warren (2022) asked participants to follow a group of textured objects, whose heading direction or speed of motion was briefly perturbed. Overall, locomotor responses were consistent with boundary motion (feature tracking). When the object texture and boundaries moved in Opposite directions (the reverse-phi illusion), responses decreased with increasing boundary blur, but this did not occur in the Same direction condition. This widening gap between the two conditions indicates that visually-guided locomotion depends on a weighted combination of feature-tracking and motion energy. Here, we leverage deep neural networks to investigate what network architectures can replicate these effects. We evaluated several model architectures, including single-stream 3D convolutional networks (MC3, R(2+1)D, R3D, DorsalNet), a dual-stream network (SlowFast), and a benchmark neurophysiological motion energy model (Nishimoto and Gallant 2011). The dual network includes a low temporal-frequency/high spatial-frequency stream, and a high temporal/low spatial stream. These models were fine-tuned to estimate the heading and speed of a group of objects with attached surface textures moving across various backgrounds from 12-frame video sequences, as in ecological contexts. The fine-tuned models were then tested on Zhu and Warren’s (2022) reverse-phi stimuli. The mean heading estimates of the single-stream and motion energy models decreased significantly in both Same and Opposite conditions as boundary blur increased (range of F(2, 594): 16.45 - 96.00, all p < 0.0001), deviating from human responses. These findings suggest that single-stream 3D convolutional networks function similarly to motion energy detectors, without the feature-tracking observed in humans. However, the SlowFast network failed to replicate the increasing gap between Same and Opposite conditions with boundary blur. We conclude that the SlowFast model does not capture human-like feature tracking. This indicates the need for further architectural improvements, such as incorporating recurrent connections to support feature-tracking.

  PublisherDOI

• An active electronic, high-density epidural paddle array for chronic spinal cord neuromodulation

  Journal of Neural Engineering · 2025-03-19 · 1 citations

  articleOpen access

  Abstract Objective . Epidural electrical stimulation (EES) has shown promise as both a clinical therapy and research tool for studying nervous system function. However, available clinical EES paddles are limited to using a small number of contacts due to the burden of wires necessary to connect each contact to the therapeutic delivery device, limiting the treatment area or density of epidural electrode arrays. We aimed to eliminate this burden using advanced on-paddle electronics. Approach . We developed a smart EES paddle with a 60-electrode programmable array, addressable using an active electronic multiplexer embedded within the electrode paddle body. The electronics are sealed in novel, ultra-low profile hermetic packaging. We conducted extensive reliability testing on the novel array, including a battery of ISO 10993-1 biocompatibility tests and determination of the hermetic package leak rate. We then evaluated the EES device in vivo , placed on the epidural surface of the ovine lumbosacral spinal cord for 15 months. Main results. The active paddle array performed nominally when implanted in sheep for over 15 months and no device-related malfunctions were observed. The onboard multiplexer enabled bespoke electrode arrangements across, and within, experimental sessions. We identified stereotyped responses to stimulation in lower extremity musculature, and examined local field potential responses to EES using high-density recording bipoles. Finally, spatial electrode encoding enabled machine learning models to accurately perform EES parameter inference for unseen stimulation electrodes, reducing the need for extensive training data in future deep models. Significance . We report the development and chronic large animal in vivo evaluation of a high-density EES paddle array containing active electronics. Our results provide a foundation for more advanced computation and processing to be integrated directly into devices implanted at the neural interface, opening new avenues for the study of nervous system function and new therapies to treat neural injury and dysfunction.

  PublisherDOI

Recent grants

• CRCNS US-France Research Proposal: Oscillatory processes for visual reasoning in deep neural networks

  NSF · $549k · 2019–2023

• CAREER: Computational mechanisms of rapid visual categorization: Models and psychophysics

  NSF · $500k · 2013–2019

Frequent coauthors

• L. Li

  3217 shared

• S. De Cecco

  Radboud University Nijmegen

  2743 shared

• T. Beau

  Consejo Nacional de Investigaciones Científicas y Técnicas

  2575 shared

• B. Trocmé

  Laboratoire AstroParticule et Cosmologie

  2526 shared

• L. Roos

  Laboratoire de Physique Nucléaire et de Hautes Énergies

  2432 shared

• S. Trincaz-Duvoid

  Laboratoire de Physique Nucléaire et de Hautes Énergies

  2431 shared

• J. Ocariz

  Université Paris Cité

  2430 shared

• M. Ridel

  Université Paris Cité

  2428 shared

Awards & honors

• PAMI Helmholtz Prize for significant impact on computer visi…
• NSF Early Career Award
• DARPA’s Young Faculty Award and Director’s Award