Intel Network on Intelligent Systems

01.07.2017 bis 01.07.2017

The lab of Prof. M. Bethge works on the interface between machine learning and computational neuroscience with a focus on computer vision and unsupervised learning/generative modeling. Among the most recent publications of the lab are novel algorithms to synthesize textures and to transfer artistic style, benchmark­breaking saliency prediction, generative image modeling, and the functional classification of retinal ganglion cells. Building upon this prior work we seek to make significant advances on the following two subjects:

1. Deep interpretable image representation, generative modeling & unsupervised learning
Convolutional layers in supervisedly trained deep neural networks provide representations that are useful across a surprisingly wide range of datasets and tasks. However, apart from very coarse aspects such as equivariance w.r.t. to location and the rough ordering w.r.t. scale along the layer hierarchy, we have very limited understanding of the structure of these representations. Using image synthesis methods, generative modeling and re­modeling approaches we seek to better understand the decision-making process in these networks. In addition, we are trying to make DNNs robust against noise and adversarial attacks, and we seek to develop unsupervised methods to learn useful image representations with minimal amount of label information.

2. Biologically informed deep neural networks for robust visual inference
We systematically compare visual inference of DNNs and humans using theory­driven psychophysical paradigms. Our first aim is to determine for which tasks humans perform systematically better or worse than DNNs. Our second aim is to model human decision making behavior in these tasks. An important component of the second aim is to adapt the network architecture of current machine vision CNNs to match the information flow in the human visual system (e.g. by accounting for foveated vision). In addition, we collaborate with the groups of P. Berens and T. Euler to inform the representations in the lower layer of DNNs by the nonlinear image transformations in the retina, and collaborate with A. Tolias (Baylor College, Houston) to compare the representations and processing in higher layers of DNNs to cortical image representations in the ventral stream of the monkey visual system.