Computer Vision News - July 2016

Robnik-Sikonja and Kononenko suggest to evaluate p(x | x\i ) in a univariate fashion, in which only one pixel is removed at a time and to simulate the absence of a feature by approximately marginalizing it out: \ ≈ \ , (2) The authors of this paper argue that the univariate approach has two main drawbacks: (1) it is very unlikely that a neural network would be easily fooled and change its prediction if a single input pixel value was changed. (2) When working with natural images, a red pixel suddenly appearing in a clear-blue sky is rather low, which again makes the approximation of equation (2) not very accurate. The authors claim that a much better approximation would be to assume that a pixel’s value depends most strongly on the pixels in some neighbourhood around it. For this they suggest to adapt the Robnik-Sikonja and Kononenko method as follows: For each pixel x in the given image the small pixel patch, denoted as x w , of size k × k × 3 (RGB image) that contains x along with and a larger patch, denoted as w , of size l × l × 3 that contains the smaller patch is defined (see figure below). Now, using a Gaussian model and a missing patch x w they sampled values for the small missing patch, conditioned on the remaining pixels which give a new approximation of p(c|x\i) as follows: ( | \ ) ≈ ( | \ ) Results To demonstrate their approach, the authors used the images from ILSVRC challenge (Russakovsky et al., 2015), a large dataset of natural images from 1000 categories, and three DCNNs: 1. AlexNet (Krizhevsky et al., 2012), 2. GoogLeNet (Szegedy et al., 2015) 3. VGGnetwork (Simonyan & Zisserman, 2014). 28 Computer Vision News Research “ The continuous increase in size and performance of DCNNs makes it harder to understand the way they operate and what are the ideal parameters to make them work ” Research