Alberto Arturo Vergani

Alberto Arturo Vergani

Post-Doc in Computational Neuroscience

Project description: Visual computations using Spatio-temporal Diffusion Kernels and Traveling Waves

Biological vision is surprisingly efficient. To take advantage of this efficiency, Deep learning and convolutional neural networks (CNNs) have recently produced great advances in artificial computer vision. However, these algorithms now face multiple challenges: learned architectures are often not interpretable, disproportionally energy greedy, and often lack the integration of contextual information that seems optimized in biological vision and human perception. It is clear from recent advances in system and computational neuroscience that nonlinear, recurrent interactions in visual cortical networks are key to this efficiency ( Tang et al., 2018; Kietzmann et al., 2019). We will use inspiration from neurophysiology and brain imaging to resolve this apparent gap between traditional CNNs and biological visual systems.

In this post-doctoral project, I will address these major limitations by focusing on specific dynamical features of cortical circuits: lateral diffusion of sensory-evoked traveling waves ( Chavane et al., 2011; Muller et al., 2018) and dynamic neuronal association fields ( Frégnac et al., 2012; Frégnac et al., 2016; Gerard-Mercier et al., 2016). Indeed, the architecture of primary visual cortex (V1), the direct target of the feedforward visual flow, contains dense local recurrent connectivity with sparse long-range connections ( Voges and Perrinet, 2012). Such connections add to the traditional convolutional kernels representing feedforward and local recurrent amplification a novel lateral interaction kernel within a single layer (across positions and channels). Less studied, but probably decisive in active vision, recurrent cortico-cortical loops add a level of distributed top-down complexity which participates to the lateral integration of sensory input and perceptual context ( Keller et al., 2019). Coupled with the continuous time dynamics of cortical circuits, this elaborate multiplexed architecture provides the conditions possible for generating information diffusion through traveling waves. Inspired by recent work in neuroscience uncovering the ubiquity of these waves during visual processing, we aim to design a self-supervised CNN that will exploit these dynamics for new applications in computer vision.

The proposed work will be organized as a collaboration between two labs (INT, Marseille and UNIC, Gif) along three tasks to be integrated in a unified model:

  1. The starting point will be to extend results of self-supervised learning that we have obtained on static, natural images ( Boutin et al., 2019) showing in a recurrent cortical-like artificial CNN architecture the emergence of interactions which phenomenologically correspond to the “association field” described at the psychophysical ( Field et al., 1993), spiking ( Li and Gilbert, 2002) and synaptic ( Gerard-Mercier et al., 2016) levels.

  2. The central aim will be to develop a dynamical version of this feedback/lateral kernel in the context of the ANR Horizontal-V1 project, linking the two labs and confronted to their recent electrophysiological data pointing to different classes of spatio-temporal diffusion and different degree of anisotropies during apparent and continuous motion.

  3. The implementation of this kernel inspired by CNN theory will be compared with a biologically realistic models of the early visual system ( Antolik et al., 2019), and simulations of the lateral diffusion kernel will be developed in collaboration with Jan Antolik, external collaborator to the ANR grant. In parallel, using tools linking neural activity to VSD imaging ( Muller et al., 2014; Chemla et al., 2019), we will analyze at a more mesocopic level the role of observed traveling waves in forming efficient representations of the visual world.

Research context

This project is funded by the French National Research Agency (ANR) under the ANR Horizontal V1 grant (coordinator Y. Frégnac) which aims at understanding the emergence of sensory predictions linking local shape attributes (orientation, contour) to global indices of movement (direction, speed, trajectory) at the earliest stage of cortical processing (primary visual cortex, i.e. V1). The cross-talk between physiological and theoretical approaches is fostered by the close collaboration with the teams of Frédéric Chavane at INT and Yves Frégnac at UNIC. The theoretical work is performed in close collaboration with Lyle Muller (Western U) and Jan Antolik (Prague). This project is primarily hosted at the Institut de Neurosciences de la Timone.

References

Education

  • Research Fellow in Neuromorphic Computing., 2019

    Middlesex University, Department of Computer Science, London, UK

  • PhD in Computer Science and Mathematics, 2018

    University of Insubria, Varese, Italy