Abstract: Modern deep neural networks have been a seismic shift in areas such as computer vision. While their performance, as measured empirically, is very significant, there is still little theoretical understanding of such models. In this talk, I will review recent research that aims at understanding deep neural network in order to clarify what they learn and how they do this. I will discuss both visualisation techniques as well as quantitative approaches that measures properties such as invariance and equivariance learned by deep neural networks. I will also discuss the problem of learning general image representations, or in short visual brains, that can transcend the typical confines of machine learning applications and be used to solve many different types of computer vision problems at once.
Bio: Andrea Vedaldi is Assistant Professor of Engineering Science at the University of Oxford. His main research interest is machine learning applied to semantic image understanding. He is author of more than 50 peer reviewed publications (h-index 37). He is the recipient of the ERC Starting Grant, the Mark Everingham Prize, the ACM Open Source Award for leading the VLFeat project, and a number of best paper awards. He is supported by the ERC, EPSRC, BP, Continental, and XRCE.