Now scientist from Italian research institute SISSA and the Technical University of Munich have found a light to shine inside – an approach for studying deep neural networks that reveals the processes that they are able to carry out – so long as they are image processing networks. "We have developed a method to systematically measure the level of complexity of the information encoded in the various layers of a deep network – the so-called intrinsic dimension of image representations," according to SISSA scientists Davide Zoccolan and Alessandro Laio. "Thanks to the collaboration of experts in physics, neurosciences and machine learning, we have exploited a tool originally developed in another area to study the functioning of deep neural networks". Working with Jakob Macke, of TUMunich, they applied the method to find out that, inside an image recognition deep neural network, representations of the image undergo a progressive transformation. Similar to what happens in the visual system, they analyse content progressively, through a chain of processing stages.