Skip to main content Skip to main navigation


Optimum statistical representation obtained from an intermediate feature level of the visual hierarchy

Noshaba Cheema; Lex Fridman; Ruth Rosenholtz; Christoph Zetzsche
In: Thomas Barkowsky; Zoe Falomir Llansola; Holger Schultheis; Jasper van de Ven (Hrsg.). Biannual Conference of the German Cognitive Science Society. Fachtagung der Gesellschaft für Kognitionswissenschaft (KogWis-2016), Space for Cognition, located at 13th, September 26-30, Bremen, Germany, Pages 63-66, Universität Bremen: Informatik/Mathematik, 9/2016.


Representations obtained from the statistical pooling of features gain increasing popularity. The common assumption is that low-level features are best suited for such a statistical pooling. Here we investigate which level of a visual feature hierarchy can actually produce the optimal statistical representation. We make use of the award-winning VGG 19 deep network which showed human-like performance in recent visual recognition benchmarks. We demonstrate that the optimum statistical representation is not obtained with the early-level features, but with those of intermediate complexity. This could provide a new perspective for models of human vision, and could be of general relevance for statistical pooling approaches in computer vision and image processing.

Weitere Links