< Back to previous page

Project

Temporal dynamics in deep neural networks: from neural mechanism to perception.

An important goal of cognitive sciences is to connect cognition with underlying neural processes. Deep neural networks (DNNs) are a modern tool for modelling complex information processing comparable to our visual system. By simulating the system at the level of neurons, DNNs model visual object recognition simultaneously in terms neural representations as well as behavioral responses. Yet, current state-of-the-art models lack the temporal dynamics of biological vision, such as recurrent connections or neural adaptation. The latter refers to the mechanisms by which our brain adjusts to the current environment based on recent sensory experience. Here, we propose to incorporate such temporal dynamics, starting with local and bottom-up mechanisms of neural adaptation. By implementing them in a fully functional neural network, we will be able to study their effect on neural representations and perception in silico. We will validate the implementation of neural adaptation with previously reported neural data from macaque visual cortex. Next, we will compare these simulations with data from psychophysical experiments and human physiological recordings. In addition, we will implement recurrent dynamics in the form of lateral and feedback connections, which will effectively increase the depth of the network. In short, for the first time we will investigate how complex perceptual and neural phenomena could emerge from simple neural mechanisms in the context of a complex network.
 

Date:1 Oct 2018 →  30 Sep 2021
Keywords:perception, deep neural networks
Disciplines:Neurosciences, Biological and physiological psychology, Cognitive science and intelligent systems, Developmental psychology and ageing