Computer Vision News - July 2016

Every month, Computer Vision News reviews a research from our field. This month we have chosen to review A New Method to Visualize Deep Neural Networks , a research paper presenting a novel visualization method that finds and highlights the regions in image space that activate the nodes (hidden and output) in the neural network. It also shows several examples of how the method can be deployed and used to understand the classifier’s decision. We are indebted to the authors (Luisa M. Zintgraf, Taco S. Cohen, Max Welling) for generously allowing us to use their images to help illustrate this review. The full paper is here . The popularity, performance and flexibility of Deep Convolutional Neural Networks (DCNNs) has vastly increased in recent years. In particular, their contribution is becoming more and more precious as classifiers able to deliver natural image classification. On the other hand, their continuous increase in size and performance makes it harder to understand the way they operate and what are the ideal parameters to make them work. Visualizing the decision-making process and inner workings of deep neural networks can greatly improve the way we train them as well as our capability to detect and correct their current weaknesses. Following the path of numerous methods previously studied to analyze deep neural networks through visualization, this research identifies and explains a novel visualization method that finds and highlights the regions in image space that activate the nodes (hidden and output) in the neural network. The author start by presenting the work of Robnik-Sikonja and Kononenko (2008) which propose a method for evaluating the predictions of probabilistic classifiers by looking at how the prediction changes if a feature (e.g. image pixel) is unknown. To express this notion they assigned each input pixel i with a weight of evidence (WE ) which is computed as follows: WE (c|x) = log (odds(c|x))-log(odds(c|x\i)) s.t. odds(c|x) = p(c|x) / (1-p(c|x) ). Where c: The image class x: The input images x\i: The input images where the i-th image pixel is treated as unknown Taking this approach the main question is how to approximate p(c|x\i) (i.e. the probability of class 'c' where the i-th image pixel is unknown) ? Once the class probability p(c│x\i) is estimated, it can be compared to p(c|x) (all image pixels are known). The underlying idea is that if there is a large prediction difference, the feature must be important. Computer Vision News Visualize Deep Neural Networks Computer Vision News Research 27 Research