Projects

Higher-Order Spatiotemporal Correlations

• Convolution goes higher-order: a biologically inspired mechanism empowers image classification - Submitted to NeurIPS 2025
  • Demonstrated that incorporating multiplicative interactions between filter outputs significantly improves performance on standard computer vision benchmarks
  • Developed a novel architecture that captures complex dependencies in visual data that standard CNNs miss
• Higher-Order Convolution Improves Neural Predictivity in the Retina - Submitted to CCN 2025
  • Extended the higher-order convolution framework to neural response prediction in the retina
  • Showed improved predictive performance while maintaining model interpretability

Information Decomposition and Generative Models

• What’s Inside Your Diffusion Model? A Score-Based Riemannian Metric to Explore the Data Manifold - Submitted to NeurIPS 2025
  • Introduced a mathematically rigorous framework for analyzing the geometry of visual representations using metrics derived from the score function in diffusion models
  • Exploited the Riemannian geometry of the learned image manifold to provide tools for understanding meaningful transformations in visual space
• Decomposing stimulus-specific sensory neural information via diffusion models - Submitted to NeurIPS 2025
  • Established an axiomatic framework for pixel-wise and feature-wise information decomposition using score functions from diffusion models

Equivariant Architectures for Visual Processing

• Scale-Equivariant Networks for Neural Response Prediction - In Progress
  • Adapted Scale-Equivariant Steerable Networks for neural response prediction to naturalistic video stimuli
  • Achieved equivalent or superior performance using only 16k parameters, compared to conventional CNNs requiring over 100k parameters (84% reduction)
• Velocity-Equivariant Video Understanding - In Progress
  • Developing a unified spatiotemporal equivariance framework that maintains parameter efficiency while enhancing performance on video understanding tasks
  • Validating preliminary results on dynamic stimuli

Spatiotemporal Gaussian Processes for Predicting Retinal Neural Responses

• Scalable gaussian process inference of neural responses to movies Accepted Poster at CoSyNe 2024
  • Developed a scalable Gaussian process framework for modeling neural responses to naturalistic movie stimuli
  • Achieved results comparable with SOTA CNN models

Functional Cell-typing in large-scale electrophysiological recordings

• Towards end-to-end cell-typing in large-scale electrophysiological recordings Accepted Poster at CoSyNe 2024
  • Created an end-to-end framework for automated identification of cell types in large-scale neural recordings
  • Leveraged functional response properties to classify different neuronal cell types
• Estimating the Surround of Ganglion Cells in Large-Scale Recordings. see Bernstein 2023 Poster
  • Characterized center-surround receptive field properties in retinal ganglion cells
  • Developed methods for accurate estimation in large-scale multi-electrode array recordings

Visual detection of elementary image features during natural behaviour

• Work led by Jun Yang at ENS, Paris. The project involved a VR experiment and training different DNNs. My contribution is essentially on the DL part. Submitted to ECVP 2025
• Used deep neural networks to predict human perceptual decisions from trial-by-trial judgments, demonstrating that both local feature analysis and global scene context contribute equally to visual decision-making in naturalistic environments