Prof. Jenny Benois Pineau
University of Bordeaux
From attention models and saliency maps to the explainability of Deep Neural Networks
Deep Neural Networks are now winner models in supervised learning for various visual content analysis/classification tasks. They proceed by building more and more abstract representations of the content for finally classify it accordingly to the user-defined taxonomy. In the lecture wi will focus on Convolutional Neural Networks which have been specifically designed to classify data organized on grids, such as images. In order to make them even more efficient, internal attention models have been recently proposed. The attention mechanisms developed are both global and local. Global attention mechanisms learn importance of feature channels in the convolutional channels and can be used for optimization of network hyper parameters, such as number of convolutional filter at each layer. Local attention mechanisms allow for re-inforcing features in the feature maps locally. We will focus first on the global, local and double attention mechanisms in this lecture. Despite the fact that the DNNs emulate cognition process, their internal « attention » derived from attention mechanisms is not « visual attention » the human deploy in the process of recognition of visual content. We will also present results of our research on the use of visual attention models to guide DNNs in their recognition tasks. All these mechanisms have built a basis for explanation of classification decisions of DNNs. This is an important and very quickly developing research area due to the necessity of interpretation of results fro decision makes. In the whole set of explanations modes which have been proposed until now, we will focus on « white box » methods which analyse futures in convolutions layers and allow for identifying regions and pixels in images which have contributed in th decision the most.