Post-Doctoral position (12 months extensible to 24 months)
Graphs are nowadays a common mathematical formalism used in various domains where the notion of network is significant, such as Genetics, Sociology, Ecology and Neurosciences for instance. For the latter graph representation allows to describe brain connectivity both at a structural and a functional level (1). Moreover, graph neural network is an emerging topic in data mining where the graph modelling allows the use of mathematical tools from graph theory in combination with deep learning approach.
The objective of this Post-Doctoral position is to use mathematical properties of classical graph neural networks in order to explore unresolved specific weaknesses of deep neural networks (DNN). We will focus on catastrophic forgetting (or catastrophic interference) and adversarial attack. The former hampers the training phase of DNN when the trained model forgets a previously learned pattern when confronted with new examples to learn. The latter refers to the vulnerability of DNN to a subtle carefully designed change in how inputs are presented completely alters its output and leads to wrong conclusion.
To study these major DNN drawbacks notably for medical application (2), we propose to represent the DNN as a graph and track learning and prediction under different conditions of training and attack (3). The final goals are to respond to several questions: are specific hidden neurons (or layers) vulnerable to forgetting or attack? Which solutions can be implemented (introduction of penalty during the training phase, specific architectures including feedback connections, …) to design DNN more resilient to forgetting and attack?
(1) Hanczar, B., Zehraoui, F., Issa, T. et al. (2020) Biological interpretation of deep neural network for phenotype prediction based on gene expression. BMC Bioinformatics 21, 501.
(2) Finlayson, S.G., Bowers, J.D., Ito, J., Zittrain, J.L., Beam, A.L., and Kohane, I.S. (2019). Adversarial attacks on medical machine learning. Science 363, 1287-1289.
(3) Bach, S., Binder, A., Montavon, G., Klauschen, F., Müller, K., & Samek, W. (2015). On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation. PLoS ONE, 10.