The goad of my research is to better understand (visual) perception. To this aim, I propose to pushfoward three directions : the first aims at providing better models of natural images and movies; the second aims at studying theoretically and numerically some key observations and hypotheses made in biological neural network; the last aims at collecting data to explore further visual perception.

  • Natural Image Models for Vision Studies

The ultimate goal is to have a controlled model of natural movies/images. First, I am interested in texture synthesis. I exploit the potential of the Stochastic Partial Differential Equation (sPDE) formulation of dynamic textures1,2 to provide well-controlled non-stationary (in time) dynamic textures to experimentalists (see some stationary examples). In particular, I explore the possibility of on the fly texture synthesis of spatially naturalistic textures3. Second, I am interested in image segmentation. I work on incorporating a key missing ingredient to the probabilistic segmentation algorithm I have developed4 : contours. My approach is to study contour statistics in natural image in order to find a generative model of natural image contours.

  • From Theory of Perception to Machine Learning

This line of research follows from ideas that I have encounter along my research path and is more prospective. First, I would like to explain the receptive field organization in the visual cortex from key hypotheses such as sparsity or the maximum information principle. Second, I would like to incorporate a normalization mechanism to artificial neural networks in the form of a feedback signal. Finally, I would like to tackle the question of optimal states in neural networks. These states, assumed to be stationary, would be of maximum information enabling efficient representation of new perceptual inputs.

  • Experimental Vision

Models must be confronted to experimental data. First, I contribute to experimental vision by improving psychometric protocols leveraging computational methods or synthesizing well controlled stimuli2. Second, I focus on detailing and testing a model of perception based on information theory. My experiments consist in testing predictions of this theory in psychophysics but bearing in mind the possible consequences for neurophysiological signals. In addition, I would like to study the possibility of using real time stimulation to develop and improve brain computer interfaces. Finally, I aim at using the new protocol I develop5 to measure human visual segmentation of images to build a large dataset of human segmented images.


  1. Vacher, J., Meso, A. I., Perrinet, L. U. & Peyré, G. Biologically inspired dynamic textures for probing motion perception. in Advances in Neural Information Processing Systems (2015).
  2. Vacher, J., Meso, A. I., Perrinet, L. U. & Peyré, G. Bayesian modeling of motion perception using dynamical stochastic textures. Neural computation 30, 3355–3392 (2018).
  3. Vacher, J., Davila, A., Kohn, A. & Coen-Cagli, R. Texture Interpolation for Probing Visual Perception. Advances in Neural Information Processing Systems (2020).
  4. Vacher, J., Launay, C. & Coen-Cagli, R. Flexibly Regularized Mixture Models and Application to Image Segmentation. Neural Networks 149, 107–123 (2022).
  5. Vacher, J., Mamassian, P. & Coen-Cagli, R. Measuring Human Probabilistic Segmentation Maps. in Cosyne Abstracts (2020).

© 2019. All rights reserved.

Powered by Hydejack v8.4.0