Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Unravelling the relationship between location and categorisation improves convolutional neural networks

Jean-Nicolas Jérémie¹, Emmanuel Daucé², Laurent Perrinet³; ¹Université d'Aix-Marseille, ²Ecole Centrale de Marseille, ³Aix Marseille Univ

Presenter: Jean-Nicolas Jérémie

Many studies have attempted to enhance the performance of convolutional neural networks (CNNs) by increasing model complexity, adding parameters, or adopting alternative architectures. Our approach differs in that we prioritise ecological plausibility in order to achieve high accuracy with minimal computational cost. We focus on visual search, which requires the localisation and categorisation of a target object in natural scenes. Due to the inhomogeneity of foveal retinotopy in human visual representations, localisation plays a key role in correctly categorising labels of interest when performing this task. We propose a framework referred to as a 'likelihood map', based on the probability of correctly identifying the target label, which explores prediction by a dedicated network according to the position of the fixation point. Depending on the scenario, it can be guided (or not guided) by the target label in a manner similar to Grad-CAM or DFF. In both scenarios, we demonstrate improved classification performance when the sensor shifts towards the region of interest. Beyond its computational benefits, this framework can be used as an experimental tool to further investigate the neural mechanisms underlying visual processing.

Topic Area: Visual Processing & Computational Vision

Extended Abstract: Full Text PDF