Elucidating the core computational principles underlying visual recognition of natural scenes
Website University of Amsterdam - Brain and Cognition
Our brains can capture the content of a photograph, picture, video in the blink of an eye. While artificial systems in recent years have improved dramatically on visual tasks like object classification, they still lack core abilities to deal with newly introduced visual conditions such as lighting changes. Human visual processing does not (or to a lesser extent) exhibit this shortcoming. It is believed that facilitation by pre-cortical structures, including the retina in the eye, leads to highly robust representations in human brains. However, current state-of-the-art models of visual processing, such as convolutional neural networks, typically lack such retinal preprocessing. In this project, we aim to investigate the contribution of these computations empirically by developing models of pre-cortical processing and testing their ability to predict human data.
To achieve this goal, we have a unique opportunity to work with an exciting new imageset. The Open Amsterdam Dataset (OADS) is a dataset of high-resolution natural outdoor scenes taken in Amsterdam created by students and researchers from the UvA. The high-resolution nature of the images enables us to better assess how the human brain responds to realistic visual inputs. In order to compare computational models, trained on the OADS dataset, not only to human behaviour (e.g., eye-tracking or response time) but also to neural responses, we plan to collect EEG data of humans viewing natural scenes from the OADS dataset. Furthermore, we aim to investigate to that extent implementing pre-cortical processes in computational models of the visual cortex aids visual recognition and if this makes computational models more similar to human brains. We expect there to be a positive influence not only on e.g., image classification performance but also our ability to understand brain responses.
This project entails: setup of a suitable experimental paradigm in an EEG data collection setting, actual collection of EEG data and the computational modelling of neural responses using state of the art techniques.
This project is intended to partly be a shared project with other master students, especially for the data collection phase. Data collection will take place at the Psychology Department of the University of Amsterdam.
Starting Date: January 2023 onwards
Duration: Minimum of 5 months
You will be supervised by Steven Scholte (UHD) and Iris Groen (UD) with supervision on a daily basis by Niklas Müller (PhD student).
To apply or ask for more information you can email Niklas Müller (firstname.lastname@example.org) and CC both Iris Groen (email@example.com) and Steven Scholte (H.S.Scholte@uva.nl).
To apply for this job email your details to firstname.lastname@example.org